Vehicle-to-Vulnerable Road Users

Channel Modeling in Critical

Scenarios

vorgelegt von
M. Sc.
Ibrahim Rashdan
ORCID: 0000-0001-9208-8078

an der Fakultät IV - Elektrotechnik und Informatik

der Technischen Universität Berlin
zur Erlangung des akademischen Grades
Doktor der Ingenieurwissenschaften
-Dr.-Ing.-
genehmigte Dissertation

Promotionsausschuss:
Vorsitzender: Prof. Slawomir Stanczak
Gutachter: Prof. Giuseppe Caire
Gutachter: Prof. Uwe-Carsten Fiebig
Gutachter: Prof. Thomas Kürner
Tag der wissenschaftlichen Aussprache: 19. Mai 2023

Berlin 2023
 Zusammenfassung

Nach Angaben der Weltgesundheitsorganisation fordern Verkehrsunfälle jedes

Jahr weltweit etwa 1,35 Millionen Menschenleben und verursachen mehr als 50
Millionen Verletzungen. Fast die Hälfte der Verkehrsopfer sind Fußgänger, Radfahrer
und Motorradfahrer. Diese Verkehrsteilnehmer sind im Straßenverkehr besonders
gefährdet, wodurch sie auch als schwache Verkehrsteilnehmer oder Vulnerable Road
Users (VRU) benannt werden. Die direkte Kommunikation zwischen Fahrzeugen
und vulnerablen Verkehrsteilnehmern (V2VRU) kann Unfälle verhindern, indem sie
eine Rundumsicht ermöglicht und die gegenseitige Erkennung und Lokalisierung von
Fahrzeugen als auch von Verkehrsteilnehmern verbessert. Ein realistischer Kanal ist
eine entscheidende Voraussetzung für die Entwicklung eines zuverlässigen V2VRU-
Kommunikationssystems. Im Gegensatz zur Fahrzeug-zu-Fahrzeug- (V2V) und
Fahrzeug-zu-Infrastruktur (V2I) Kommunikation wurde der V2VRU-Kommunikation
in der Forschung noch nicht viel Aufmerksamkeit geschenkt. Ein dediziertes
Kanalmodell für V2VRU-Kommunikation in kritischen Unfallszenarien ist noch
nicht vorhanden. Um hier Abhilfe zu schaffen, zielt diese Arbeit darauf ab, die erste
vollständige Parametrisierung eines geometriebasierten stochastischen Kanalmodells
(geometry-based stochastik channel model, GSCM) für kritische Unfallszenarien in
Städten zu erstellen. Zu diesem Zweck wurden experimentelle Single-Input-Single-
Output (SISO)-Kanalmessungen sowohl im freien Feld als auch in städtischen
Umgebungen durchgeführt. Für die Kanalmessungen wurden Signale mit einer
Bandbreite von 120 MHz bei einer Trägerfrequenz von 5,2 GHz eingesetzt, die in
der Nähe des 5,9-GHz-ITS-G5-Bandes und des 5,7-GHz-ISM-Bandes (Industrie,
Wissenschaft und Medizin) liegen. Dabei wurden kritische Unfallszenarien mit
Fahrzeugen und VRUs in den Messungen nachgestellt.
 Obwohl sich eine Handvoll neuerer Studien mit dem Pfadverlust des Fahrzeug-
Fußgänger-Kanals (V2P) befasst haben, ist nur wenig über die Auswirkungen der
Mobilität des Fußgängers, der Behinderung durch geparkte Fahrzeuge und der
Abschattung durch eine umgebende Menschenmenge auf die Empfangsleistung
bekannt. In dieser Arbeit werden diese Aspekte untersucht und neue Modelle für
den Pfadverlust vorgeschlagen. Außerdem wird der Beugungsverlust aufgrund der
Behinderung durch geparkte Fahrzeuge berechnet und modelliert. Die Erkenntnisse
über den Beugungsverlust werden außerdem durch Simulationen unterstützt, die
zeigen, dass das Multiple-Knife-Edge Modell eine gute Übereinstimmung mit dem
gemessenen Beugungsverlust bietet.
Es ist in der Literatur gut belegt, dass Fahrzeugkanäle nicht stationär sind.
Um ein GSCM-Kanalmodell zu parametrisieren, ist daher die Kenntnis der
Stationaritätsdistanz erforderlich. Die Nicht-Stationarität des V2VRU-Kanals wurde
jedoch in der Literatur noch nicht analysiert. Daher wird in dieser Arbeit die
Nicht-Stationarität des V2VRU-Kanals untersucht und die Stationaritätsdistanz
geschätzt. Darüber hinaus wird festgestellt, dass die zeitvariante Kanalimpulsantwort
(Channel Impulse Response, CIR) in der städtischen Umgebung durch diffuse
Mehrwegekomponenten (Diffuse Multipath Components, DMCs) stark geprägt
ist. Um eine Charakterisierung der reflektierte Mehrweg Komponenten (Specular
Multipath Components, SMCs) zu ermöglichen, wird eine neuartige Methode zur
Extraktion der SMCs aus der CIR basierend auf der Dichte ihrer benachbarten
Mehrwegkomponenten vorgeschlagen. Darüber hinaus wird ein Algorithmus zur
Verfolgung von SMCs über die Zeit auf der Grundlage ihrer Verzögerung und
Amplitude vorgestellt. Um einen besseren Einblick in die Entwicklung des Funkkanals
zu erhalten, wird die Position der Streuer in der Ausbreitungsumgebung mit Hilfe
eines Algorithmus zur gemeinsamen Verzögerungs-und-Doppler-Schätzung ermittelt.
Schließlich wird in dieser Arbeit eine vollständige Parametrisierung für das
GSCM vom WINNER-Typ vorgeschlagen. Insbesondere werden die Kanalparameter
und ihre (Kreuz-) Korrelationen im log-Bereich geschätzt. Die Ergebnisse zeigen,
dass die lognormale Verteilung eine gute Annäherung an die Verteilungen der
Kanalparameter bietet. Nach der Parametrisierung werden Kanalsimulationen
mit dem quasi deterministischen Funkkanalgenerator (QuaDRiGa) durchgeführt.

iv
Abschließend wird der GSCM mit den vorgeschlagenen Parametern validiert. Die
Kanalvalidierung zeigt, dass das vorgeschlagene Modell eine sehr gute Darstellung
des V2VRU-Ausbreitungskanals in den betrachteten Szenarien liefert. Daher kann
das vorgeschlagene Kanalmodell in Simulationen verwendet werden, um V2VRU-
Kommunikation und Kollisionsvermeidungsalgorithmen in kritischen Unfallszenarien
zu entwickeln und zu bewerten.

v
 Abstract

According to the world health organization, traffic accidents take about 1.35 million
lives and cause more than 50 million injured persons globally each year. Vulnerable
road users (VRUs), i.e., pedestrians, cyclists and motorcyclists, account for almost
half of the road victims. Direct vehicle-to-VRU (V2VRU) communication can prevent
accidents by providing 360◦ awareness and improving detection, localization, and
tracking of both vehicles and VRUs. Having a realistic channel is a prerequisite for
developing a reliable V2VRU communication system. Contrary to vehicle-to-vehicle
(V2V) and vehicle-to-infrastructure (V2I) communications, V2VRU communication
did not attract much attention in research. A dedicated channel model for V2VRU
communication in critical accident scenarios is still missing. In order to remedy this
situation, this thesis aims to provide the first full parametrization for a geometry-
based stochastic channel model (GSCM) for critical urban scenarios. For this
purpose, experimental single-input single-output (SISO) channel measurements
were conducted in both open-field and urban environments. The measurements
were carried out at a carrier frequency of 5.2 GHz which is close to the 5.9 GHz
ITS-G5 band and to the 5.7 GHz industrial, scientific and medical (ISM) band. The
measurements were executed with a bandwidth of 120 MHz taking into account the
most critical accident scenarios involving vehicle and VRUs.
Even though a handful of recent studies addressed the path loss of the vehicle-
to-pedestrian (V2P) channel, little is known about the impact of the pedestrian
mobility, obstruction by parked vehicles, and shadowing by a crowd surrounding the
pedestrian on the received power. In this thesis, these aspects are investigated and
path loss models are proposed. Moreover, the diffraction loss due to the obstruction
of parked vehicles is calculated. The findings on the diffraction loss are then verified
by simulations. It is shown that the multiple knife-edge model provides a good
match to the measured diffraction loss.
Note that it is well established in literature that vehicular channels are non-
stationary. Therefore, in order to parameterize a GSCM channel model, the
stationarity distance is required. However, the non-stationarity of the V2VRU
channel has not yet been analyzed in literature. Hence, in this work, the non-
stationarity of the V2VRU channel is investigated and the stationarity distance is
estimated. Furthermore, the time-variant channel impulse response (CIR) in the
urban environment is found to be highly cluttered by diffuse multipath components
(DMCs). To allow for further characterization of the specular multipath components
(SMCs), a novel method is proposed to extract the SMCs from the CIR based on the
density of their neighboring multipath components (MPCs). Further, an algorithm
for tracking SMCs over time based on their delay and magnitude is presented. In
order to gain more insight on the evolution of the radio channel, the locations of
all scatterers in the propagation environment are estimated by employing a joint
delay-Doppler estimation algorithm.
Finally, the thesis proposes a full parametrization for the WINNER-type GSCM.
In particular, the large scale parameters (LSPs) and their correlations are estimated
in the log domain. The results show that the log-normal distribution provides a
good fit to the distributions of the LSPs. Following the parameterization, channel
simulations are performed with the quasi deterministic radio channel generator
(QuaDRiGa) implementation. Thereafter, the GSCM with the proposed parameters
is validated. The channel validation shows that the proposed model provides a very
good representation for the V2VRU propagation channel in the considered scenarios.
The proposed channel model can be used in simulations to develop and evaluate
V2VRU communication and collision avoidance algorithms in critical accident
scenarios.

viii
List of Published Content
The thesis is in part based on seven original publications. The author has had the
main responsibility for performing the analysis and writing all the papers.

1. I. Rashdan, S. Sand, S. Jiang, W. Wang, and G. Caire. “Non-Stationarity

Analysis of Vehicle-to-Vulnerable Road Users Channel in Critical Scenarios”.
In: 2023 17th European Conference on Antennas and Propagation (EuCAP),
2023.

2. I. Rashdan, P. Unterhuber, F. de Ponte Müller, S. Sand, and G. Caire

“Diffuse Multipath Analysis of Vehicle-to-Vulnerable Road User Channel in
Urban Environment”. In: 2021 15th European Conference on Antennas and
Propagation (EuCAP), 2021.

3. I. Rashdan, F. de Ponte Müller, S. Sand, T. Jost, and G. Caire “Measurement-

based Geometrical Characterisation of the Vehicle-to-Vulnerable-Road-User
Communication Channel”. In: IET Microwaves, Antennas and Propagation,
2020.

4. I. Rashdan, M. Walter, W. Wang, and G. Cair “Large Scale Fading

Characteristics for Vehicle-to-Cyclist Channel in Urban Environment at 5
GHz”. In: 2020 14th European Conference on Antennas and Propagation
(EuCAP), 2020.

5. I. Rashdan, F. de Ponte Müller, T. Jost, and S. Sand “Measurement-based

Geometrical Characterization of the Vehicle-to-Pedestrian Channel”. In: 2019
13th European Conference on Antennas and Propagation (EuCAP), 2019.

6. I. Rashdan, F. de Ponte Müller, T. Jost, S. Sand, and G. Caire “Large-

Scale Fading Characteristics and Models for Vehicle-to-Pedestrian Channel at
5-GHz”. In: IEEE Access, 2019.

7. I. Rashdan, F. de Ponte Müller, W. Wang, M. Schmidhammer, and S. Sand

“Vehicle-to-Pedestrian Channel Characterization: Wideband Measurement

ix
Campaign and First Results”. In: IEEE 12th European Conference on Antennas
and Propagation (EuCAP). London, UK, Apr. 2018.

x
To my parents
 Acknowledgements

First and foremost, I wish to express my deepest gratitude to my supervisor Prof.

Giuseppe Caire for giving me the opportunity to pursue my Ph.D. studies under his
guidance. I am very grateful for the help of Dr. Stephan Sand, my co-supervisor Dr.
Fabian de Ponte Müller, and Prof. Uwe-Carsten Fiebig for their invaluable guidance,
encouragement, and support. Furthermore, I am thankful to Prof. Thomas Kürner
for his interest and the examination of this thesis.
Thanks to all my colleagues for all the fruitful discussions we had, and their
support in conducting the measurement campaigns. I am fortunate to work in such
friendly environment that you created.
I am especially grateful to my dear friends, Adnan, Ammar, Dina, Nayef, Soli,
and Yazan. Thanks for sharing all the happiness and sadness. I feel so lucky to have
you in my life. Above all, I would like to thank my family; my parents and siblings.
I would never have made it this far without your unconditional support and love.

Ibrahim Rashdan

Gilching, July 2022

 Table of Contents

Title Page i

Zusammenfassung iii

Abstract vii

List of Figures xix

List of Tables xxv

Acronyms xxvii

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Contribution and Structure of this Thesis . . . . . . . . . . . . . . . 4

2 Wave Propagation and Channel Modeling Fundamentals 9

2.1 Wave Propagation Mechanisms . . . . . . . . . . . . . . . . . . . . 9
2.2 Wireless Channel Models . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Deterministic Channel Models . . . . . . . . . . . . . . . . . 11
2.2.2 Stochastic Channel Models . . . . . . . . . . . . . . . . . . . 12
2.2.3 Geometry-based Stochastic Channel Models . . . . . . . . . 12
2.3 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Channel Model Description . . . . . . . . . . . . . . . . . . . . . . 15
2.4.1 Model Parameters . . . . . . . . . . . . . . . . . . . . . . . 15

xv
TABLE OF CONTENTS

2.4.2 Generation of the Channel Coefficients with QuaDRiGa . . . 17

3 Wideband Channel Measurements 29

3.1 Open-field Measurements . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.1 Measurement Systems . . . . . . . . . . . . . . . . . . . . . 30
3.1.2 Measurement Scenarios . . . . . . . . . . . . . . . . . . . . . 32
3.2 Urban Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2.1 Measurement Systems . . . . . . . . . . . . . . . . . . . . . 36
3.2.2 Measurement Scenarios . . . . . . . . . . . . . . . . . . . . . 36

4 Channel Characteristics and General Modeling Aspects 39

4.1 Path Loss Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2 Spatial Correlation of Shadow Fading in Open-field . . . . . . . . . 48
4.3 Diffraction Loss Model . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.4 Non-stationarity Analysis . . . . . . . . . . . . . . . . . . . . . . . 60
4.5 MPCs Parameters Estimation and Tracking . . . . . . . . . . . . . 71
4.5.1 Specular Multipath Extraction . . . . . . . . . . . . . . . . . 73
4.5.2 Specular Multipath Tracking . . . . . . . . . . . . . . . . . . 78
4.6 Scatterers Localization . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.7 Diffuse Multipath Components . . . . . . . . . . . . . . . . . . . . 92
4.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5 Channel Model Parameterization 99

5.1 Large Scale Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.1.1 Shadow Fading . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.1.2 Delay Spread . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.1.3 Narrowband K-factor . . . . . . . . . . . . . . . . . . . . . . 104
5.1.4 Angular Spread . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.2 Correlation Distances . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.3 Cross-correlation Parameters . . . . . . . . . . . . . . . . . . . . . . 112
5.4 Scaling Coefficients and Number of MPCs . . . . . . . . . . . . . . 114
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

xvi
 TABLE OF CONTENTS

6 Model Validation 117

6.1 Simulated Channels and Path Loss . . . . . . . . . . . . . . . . . . 118
6.2 Large Scale Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.2.1 Shadow Fading . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.2.2 Delay Spread . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.2.3 Narrowband K-factor . . . . . . . . . . . . . . . . . . . . . . 125
6.2.4 Angular Spread . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.3 Correlation Distances . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.4 Cross-correlation Parameters . . . . . . . . . . . . . . . . . . . . . . 132
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

7 Summary and Future Work 135

Bibliography 139

xvii
 List of Figures

1.1 Illustrations of the most critical accident scenarios in urban environ-

ments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Illustration of the main contributions and the flow of the work that
has been followed in this thesis based on the open-field measurements.
The number near the top-left corner of each box indicates the number
of the section or chapter in which the topic is introduced. . . . . . . 6
1.3 Illustration of the main contributions and the flow of the work that
has been followed in this thesis based on the urban measurements.
The number near the top-left corner of each box indicates the number
of the section or chapter in which the topic is introduced. The output
of the dashed box contains the model parameters, which are used as
an input for the simulation, and as validation metrics. . . . . . . . . 7

2.1 Steps for calculation of time-evolving channel coefficients. . . . . . . 18

2.2 Principle of the generation of correlated LSPs maps. . . . . . . . . . 18
2.3 Examples of uncorrelated and spatially correlated Gaussian shadow
fading maps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Illustration of the calculation of the scatterer positions and updates
of the departure and arrival angles in the multi-bounce model. . . . 23
2.5 Illustration of the overlapping area used for calculating the transitions
between segments. . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

xix
LIST OF FIGURES

3.1 Aerial view of the measurement scenario showing the trajectories of

the TX and the RX towards the imaginary collision point. (Google
Maps 2017 Geobasis-de/BKG.) . . . . . . . . . . . . . . . . . . . . 30
3.2 Transmitter (Tx) and receiver (Rx) involved in the open-field
measurement campaign. . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3 Four measurement scenarios in the open-field environment addressed
in this thesis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Aerial view of the measurement scenarios in the urban environment
showing the trajectory of the vehicle (in blue), and the trajectory
of the cyclist/pedestrian (in green with the number of the scenario)
towards the imaginary collision point. (Google Earth 2018 Geobasis-
DE/BKG.). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5 Antennas positions on the pedestrian, the cyclist, and the test vehicle. 36
3.6 Illustrations of the accident scenarios in urban environment. The
data collected in these scenarios are used to model the V2VRU channel. 38

4.1 The measured and modeled path loss in the open-field environment. 43
4.2 The measured and modeled path loss in the urban environment. . . 46
4.3 An example of the extraction of the SF from the measured path loss
for Scenario 4 in the open-field using a window size of 10λ. . . . . . 49
4.4 Mean square value of the SF calculated within intervals of 5 m and
10 m for Scenario 1 and Scenario 4 in open-field. . . . . . . . . . . . 50
4.5 Score of the MRA stationarity test. . . . . . . . . . . . . . . . . . . 51
4.6 Normalized empirical spatial autocorrelation function of the SF for
Scenario 1, Scenario 4 and the two parts of Scenario 4 in open-field. 52
4.7 Normalized empirical spatial autocorrelation function of the SF and
the corresponding fitting models for Scenario 1 and Scenario 4 in
open-field. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.8 Normalized empirical spatial autocorrelation function of the SF and
the corresponding fitting models for the LoS and OLoS parts of
Scenario 4 in open-field. . . . . . . . . . . . . . . . . . . . . . . . . 54

xx
 LIST OF FIGURES

4.9 The Epstein-Peterson method for four edges. For the edge B, the
geometrical parameters (d1 , d2 , h) in Equation (4.21) are (d1 , d2 ,
hB ). Similarly, for edge C, the geometrical parameters (d1 , d2 , h) in
Equation (4.21) are (d2 , d3 , hC ). This figure can be seen as a 2D
projection of the 3D model in Figure 4.11. . . . . . . . . . . . . . . 56
4.10 Snapshot from the 3D modeling too where the Tx (black vehicle) is
shown at different positions, and the Rx (Pedestrian) is shadowed by
parked vehicles. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.11 3D illustration of the modeling procedure. . . . . . . . . . . . . . . 58
4.12 The measured path loss during obstruction of the LoS, and the
modeled path loss as a summation of the multiple knife-edge
diffraction model and the two-ray path loss model for Scenario 4
in open-field. The measured path loss is plotted in different colors
where each color corresponds to the obstruction of the LoS by a specific
vehicle. For example, the path loss in red contains the diffraction loss
caused by the vehicle near the collision point. . . . . . . . . . . . . 59
4.13 Examples of the estimated GLSF at different time instants for Scenario
3 in the urban environment. . . . . . . . . . . . . . . . . . . . . . . 63
4.14 Examples of collinearity of the GLSF in the urban environment. . . 65
4.15 Stationarity distances in urban environment. . . . . . . . . . . . . . 67
4.16 CDFs of the stationarity distances for all scenarios in the urban
environment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.17 Stationarity distance relation with the power of the LoS component
and the KF in Scenario 3. . . . . . . . . . . . . . . . . . . . . . . . 70
4.18 Examples of the time-variant CIRs based on the measured data. . . 72
4.19 Time-variant CIRs based on the estimated parameters. The geometric
LoS is displayed as a red dashed line. . . . . . . . . . . . . . . . . . 73
4.20 Example of applying the specular multipath extraction method. . . 76
4.21 The extracted specular MPCs using different values for ζ∆τ while
ζL = 10 MPCs and δ = 25 snapshots. . . . . . . . . . . . . . . . . . 77
4.22 The extracted specular MPCs using different values for ζL while
ζ∆τ =1.5 ns and δ = 25 snapshots. . . . . . . . . . . . . . . . . . . . 77

xxi
LIST OF FIGURES

4.23 Results of the specular MPCs extraction method. . . . . . . . . . . 77

4.24 Tracked path over time in the open-field environment as a results of
the KEST algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.25 The automated tuning of the delay and magnitude threshold for the
LoS component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.26 Tracked path over time for Scenario 1 in the urban environment as a
results of the proposed tracking algorithm. . . . . . . . . . . . . . . 85
4.27 Tracked paths over time in the open-field environment as a results of
KEST algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.28 Estimated Locations of the scatterers in open-field scenario. The
results not only show the true locations but also the ambiguities. The
trajectories of the Tx and the Rx are shown © Google. . . . . . . . 88
4.29 Photos of the localized scatterers in the open-field environment. . . 88
4.30 Tracked path over time for Scenario 1 in the urban environment as a
results of the proposed tracking algorithm. . . . . . . . . . . . . . . 89
4.31 Estimated Locations of the scatterers in the urban Scenario 1. The
results not only show the true locations but also the ambiguities. The
trajectories of the Tx and the Rx are shown © Google. At each time
instant, the trajectories of the Tx and Rx together with the point in
the center of the location PDF have the same color. . . . . . . . . . 90
4.32 Photos of the localized scatterers in the urban Scenario 1. . . . . . . 91
4.33 Total received power and power carried by the SMC and DMC as
a function of Tx-Rx distance for the three scenarios in the urban
environment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.34 The fractional DMC power f̂ DMC as a function of Tx-Rx distance
calculated by Equation (4.60) for the three scenarios in the urban
environment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5.1 CDFs of measured SF and the log-normal fit for the three scenarios
in urban environment. . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.2 CDFs of measured DS and the log-normal fit for the three scenarios
in urban environment. . . . . . . . . . . . . . . . . . . . . . . . . . 103

xxii
 LIST OF FIGURES

5.3 CDFs of measured KF and the log-normal fit for the three scenarios
in urban environment. . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.4 CDFs of measured ASD and the log-normal fit for the three scenarios
in urban environment. . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.5 CDFs of measured ASA and the log-normal fit for the three scenarios
in urban environment. . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.6 Example of the normalized autocorrelation function of the KF in
Scenario 3 for both LoS and OLoS parts. The black dashed line marks
the value of e−1 at which the correlation distance is calculated. . . . 110

6.1 Channel model validation process. . . . . . . . . . . . . . . . . . . . 118

6.2 Examples of measured and simulated channel impulse responses for
the three scenarios in urban environment. . . . . . . . . . . . . . . . 120
6.3 Comparison of the path loss for the three scenarios in urban environment.121
6.4 Comparison of the distribution of the shadow fading for the three
scenarios in urban environment. . . . . . . . . . . . . . . . . . . . . 123
6.5 Comparison of the distribution of the RMS delay spread for the three
scenarios in urban environment. . . . . . . . . . . . . . . . . . . . . 124
6.6 Comparison of the distribution of the narrowband K-factor for the
three scenarios in urban environment. . . . . . . . . . . . . . . . . . 126
6.7 Comparison of the distribution of the azimuth spread of departure
for the three scenarios in urban environment. . . . . . . . . . . . . . 127
6.8 Comparison of the distribution of the azimuth spread of arrival for
the three scenarios in urban environment. . . . . . . . . . . . . . . . 128

xxiii
 List of Tables

3.1 Channel sounder parameters in the open-field measurements. . . . . 31

3.2 Channel sounder parameters in the urban measurements. . . . . . . 37

4.1 Log-distance path loss model parameters in the open-field environment. 45

4.2 Log-distance path loss model parameters for urban scenarios. . . . . 47
4.3 Spatial correlation models’ parameters. . . . . . . . . . . . . . . . . 54
4.4 Average stationarity distances in the urban environment. . . . . . . 66
4.5 Specular reflections extraction parameters. . . . . . . . . . . . . . . 75
4.6 Tracking parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5.1 Large scale parameter distributions and correlation distances. . . . 111

5.2 Estimated cross-correlation coefficients. . . . . . . . . . . . . . . . 113
5.3 Estimated scaling factors in all scenarios. . . . . . . . . . . . . . . . 114

6.1 Comparison of LSPs for Scenario 1. . . . . . . . . . . . . . . . . . . 129

6.2 Comparison of LSPs for Scenario 2. . . . . . . . . . . . . . . . . . . 130
6.3 Comparison of LSPs for Scenario 3. . . . . . . . . . . . . . . . . . . 131
6.4 Comparison of cross-correlation coefficients for the three scenarios in
urban environment. . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

xxv
 Acronyms

AoA Angle of arrival

AoD Angle of departure
APDP Average power delay profile
AS Angular spread
ASA Azimuth spread of arrival
ASD Azimuth spread of departure

C-V2X Cellular-vehicle-to-everything
CDF Cumulative distribution function
CIR Channel impulse response
COST European cooperation in science and technology

DMC Diffuse multipath component

DPSS Discrete prolate spheroidal sequence
DS RMS delay spread
DSRC Dedicated short range communication

FBS First-bounce scatterer

GDP Gross domestic product

GLoS Geometric LoS
GLSF Generalized local scattering function
GO Geometrical optics

xxvii
Acronyms

GSCM Geometry-based stochastic channel model

KEST Kalman enhanced super resolution tracking

KF Narrowband K-factor

LBS Last-bounce scatterer

LoS Line of sight
LQS Local quasi-stationarity
LSF Local scattering function
LSP Large scale parameter
LTE Long term evolution

MPC Multipath component

MRA Modified reverse arrangement test

OLoS Obstructed-line of sight

PDP Power delay profile

PDR Packet delivery ratio

RMS Root-mean-square
RSSI Received signal strength indication
RT Ray-tracing
Rx Receiver

SAGE Space-alternating generalized expectation-maximization

SCM-E Spatial channel model-extended
SF Shadow fading
SISO Single-input single-output
SMC Specular multipath component

TDL Tapped-delay line

xxviii
 Acronyms

Tx Transmitter

US Uncorrelated scattering

V2I Vehicle-to-infrastructure
V2P Vehicle-to-pedestrian
V2V Vehicle-to-vehicle
V2VRU Vehicle-to-vulnerable road user
V2X Vehicle-to-everything
VRU Vulnerable road user

WINNER Wireless world initiative for new radio

WSS Wide-sense stationary
WSSUS Wide-sense stationary uncorrelated scattering

xxix
 Introduction
1
1.1 Motivation
In recent decades, road traffic has increased enormously due to the rapid growth
of population and cities. One major drawback of this expansion is the increase of
road traffic accidents. According to the World Health Organization, traffic accidents
take about 1.35 million lives and result in more than 50 million injured persons
globally each year with an associated governmental cost of about 3 % of the gross
domestic product (GDP) [1]. Vulnerable road users (VRUs), i.e., pedestrians, cyclists
and motorcyclists, account for almost half of the road victims. Based on accident
statistics reported in [2] and [3], the most critical pre-crash scenarios that involve
pedestrians and cyclists occur in urban environments. In these scenarios, shown
in Figure 1.1, the driver visibility toward the pedestrian and cyclist is blocked by
buildings at intersections or by parked vehicles along the roadside.
Currently, driver assistance systems and automated vehicles only rely on their
own perception sensors to detect and locate other surrounding traffic participants.
However, radar sensors, laser-scanners, and camera-based systems have one critical
limitation: They require a direct line-of-sight (LoS) towards the other road users.
Additionally, light-based systems show a low performance under adverse weather

1
1. Introduction

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 1.1: Illustrations of the most critical accident scenarios in urban environments.

or lighting conditions [4, 5, 6]. One way to overcome this limitation and obtain
360 degree of awareness is to use a communication technology to directly exchange
information between vehicles and VRUs [7, 8]. Using vehicle-to-VRU (V2VRU)
communication, also called vehicle-to-pedestrian (V2P) communication, can improve
mutual detection, localization, and tracking of both vehicles and VRUs. Each
vehicle and VRU periodically transmits its position and heading. Using this
information, a collision avoidance system can trigger warning messages when a
potential collision is detected. V2VRU communication is part of the vehicle-
to-everything (V2X) communication which also includes vehicle-to-vehicle (V2V)
and vehicle-to-infrastructure (V2I) communications. V2X communication has the
potential to increase safety and efficiency of transportation systems. There are
two main technologies that enable the V2X communication, i.e., WiFi-based and
cellular-based technologies. Dedicated short range communication (DSRC) in the
US or ITS-G5 in Europe are IEEE 802.11p-based technologies, which were developed
over a decade ago. The 3GPP consortium is incorporating V2X communication
capabilities into their long term evolution (LTE) standard under the name LTE-V2X
or cellular-V2X (C-V2X) communication. The reliability of the safety application

2
 1.1 Motivation

is highly dependent on the quality of the communication link between vehicles

and VRUs, which relies on the properties of the propagation channel. Therefore,
profound knowledge of the propagation channel is a prerequisite for the development
of the communication system. Hence, in order to develop a reliable communication
system and evaluate its performance, realistic channel models are required.
Several channel models were proposed for V2V and V2I communications [9,
10, 11]. However, different from V2V and V2I, V2VRU does not attract similar
attention in research. To the best of the author’s knowledge, a dedicated channel
model for V2VRU communication in realistic and critical accident scenarios in urban
environments is still missing.
Since the vehicle is a common element in V2V, V2I and V2VRU, some similarities
potentially arise in the propagation channel. Assuming that a safety system based on
V2VRU communication is incorporated in the pedestrian’s or cyclist’s smartphone,
important differences in the propagation channel can be identified due to:

• The mobility pattern and velocity of the VRU

• The changing antenna height and orientation depending on the smartphone’s

location and VRU’s activity (texting, phoning, etc)

• The relatively low height of the VRU’s smartphone antenna, such that the LoS
could be partially or completely obstructed by road side objects, e.g. trees,
moving or parked vehicles and surrounding VRUs.

These aspects will impose different propagation characteristics. Signals may

experience additional attenuation and thus they need to be accounted for when
developing the channel model. To derive an accurate model, the V2VRU channel
has to be thoroughly investigated. These channel models should be able to
reproduce reliably and with low complexity the time-variant behavior of the channel
characteristics. Additionally, due to the movement of the transmitter (Tx), the
receiver (Rx) and the scatterers, the relatively low antenna heights and the resulting
variability of the multipath richness of the environment, the V2VRU channel is
non-stationary [12]. This non-stationarity needs to be considered which makes
characterizing and modeling the channel more difficult and challenging.

3
1. Introduction

Channel models can be broadly divided in three categories: deterministic,

stochastic, and geometry-based stochastic channel models (GSCMs). The two
main factors that influence the decision of which modeling approach to follow are
the complexity and the accuracy of the model. Taking the non-stationarity nature
of the V2VRU channel into account, a deterministic approach can provide a highly
accurate and realistic channel model. However, it is site-specific and requires intensive
and time consuming computations. Stochastic models, on the other hand, have
relatively low complexity which makes them easy to use. Yet, they are inadequate to
produce realistic models for non-stationary channels [13]. Combining the geometrical
and statistical elements of the deterministic and stochastic models, GSCMs provide
accurate geometrical relations with a reduced computational cost. Moreover, GSCMs
are found to be well suited for non-stationary environments [12]. Therefore, we have
chosen the GSCM approach in modeling the V2VRU channel in this thesis. The
two main types of GSCMs are the COST 2100 which has been defined within the
European cooperation in science and technology [14] and the wireless world initiative
for new radio (WINNER II) [15]. Since the COST 2100 channel model, on the one
hand, is cluster-centric, it requires to draw the distribution of the scattering clusters
from the measurement data. Unfortunately, our collected data are based on single-
input single-output (SISO) measurements, which limits the ability to accurately
acquire such distributions. On the other hand, the WINNER-type channel model
is user-centric, i.e., the placement of the scatterers in the simulated propagation
environment is based on the large scale parameters (LSPs). These LSPs can be
estimated based on the measurement data. Therefore, the WINNER-type channel
model is chosen in this work. More details about channel modeling approaches are
presented in Section 2.2.

1.2 Contribution and Structure of this Thesis

This thesis aims at characterizing and modeling the V2VRU channel based on
channel sounding measurements in critical accident scenarios. During the work on
this thesis, the author published, as first author, two journal papers [16, 17] and
eight conference papers [18, 19, 20, 21, 22, 23, 24, 25]. Four of the conference papers

4
 1.2 Contribution and Structure of this Thesis

[18, 19, 20, 21] considered several aspects related to V2X communications, and are
not directly related to the topic of this thesis. For reasons of consistency, this work
is not included in this thesis.
The following paragraphs provide an overview of the structure and contribution
of this thesis.
Chapter 2 gives an overview of the propagation channel fundamentals by
first describing the different electromagnetic wave propagation mechanisms. Next,
different approaches for channel modeling are briefly discussed, and the recent
developments in V2VRU channel modeling are presented. Finally, a description of
the proposed channel model is provided, followed by a detailed explanation of the
channel coefficient generation procedure.
Chapter 3 provides a detailed description of the two SISO channel measurement
campaigns. Both campaigns were conducted using the RUSK-DLR channel sounder
at a carrier frequency of fc = 5.2 GHz and with a bandwidth of B = 120 MHz. The
first campaign presented in Section 3.1 was conducted in an open-field environment
considering an accident scenario between a vehicle and a pedestrian [22]. This
location was chosen since it represents a controlled environment with only a small
number of far-located scatterers. Therefore, the location makes it possible to isolate
and study the impact on the propagation channel caused by the different elements
in the propagation environment, as well as by the mobility of the Tx and Rx. In
Section 3.2, the second measurement campaign that was conducted in an urban
environment is described. The three most critical accident scenarios involving
pedestrians and cyclists were considered. The proposed channel model in this thesis
is based on data collected during this campaign.
To make the contributions clearer, Figures 1.2 and 1.3 illustrate the main
contributions and the flow of the work that has been followed in this thesis in
open-field and urban environments.
Chapter 4 addresses several aspects of channel modeling. In Section 4.1, path
loss models for open-field and urban channel are proposed [16, 23]. The path loss does
not only provide valuable insight into the impact of the propagation environment on
the received power, but it is also used as an input to the channel model. Moreover, the
spatial correlation of the shadow fading can lead to degradation in the communication

5
1. Introduction

Figure 1.2: Illustration of the main contributions and the flow of the work that has
been followed in this thesis based on the open-field measurements. The number near
the top-left corner of each box indicates the number of the section or chapter in which
the topic is introduced.

performance. Therefore, the spatial correlation of the shadow fading in open-field

is analyzed and models are proposed [16]. To study the propagation loss due to
the obstruction of the LoS by parked vehicles, a 3D tool has been developed in
Section 4.3. The tool detects the diffraction edges and calculates the Fresnel-Kirchoff
parameter that is used to calculate the knife-edge diffraction loss.
Due to the non-stationarity of the V2VRU channel, the channel is characterized
by dividing it into regions where the wide-sense stationary uncorrelated scattering
(WSSUS) assumption holds and then the LSPs are estimated in each individual
region. In Section 4.4, the non-stationarity of the V2VRU channel is analyzed by
estimating the generalized local scattering function (GLSF) and its collinearity based
on the channel measurement data in the urban environment, and the estimated
stationarity distance is presented. In Section 4.5, the multipath parameters are
estimated using the Kalman enhanced super resolution tracking (KEST) algorithm.
To separate the specular multipath components (SMCs) and the diffuse multipath
components (DMCs), a novel method is proposed to extract specular reflections from
the estimated time-variant channel impulse responce (CIR) based on the density of
their neighboring multipath components (MPCs). This extraction allows for further

6
 1.2 Contribution and Structure of this Thesis

characterization of the specular reflections. Furthermore, a simple but effective

algorithm for multipath tracking based on the differences in delay and magnitude
between SMCs is presented [17]. Based on the previous step, in Section 4.6, the
SMC parameters are employed to localize the physical scatterers in the propagation
environment using a joint delay-Doppler estimation algorithm [17, 24]. The estimated
positions of the scatterers are then used to estimate the angle of departure (AoD)
and angle of arrival (AoA) of the SMCs. In Section 4.7, the DMCs are extracted in
order to calculate their contribution to the total received power [25].

Figure 1.3: Illustration of the main contributions and the flow of the work that has
been followed in this thesis based on the urban measurements. The number near the
top-left corner of each box indicates the number of the section or chapter in which
the topic is introduced. The output of the dashed box contains the model parameters,
which are used as an input for the simulation, and as validation metrics.

7
1. Introduction

Chapter 5 explains how the LSPs are estimated and modeled. For developing
a WINNER-type GSCM for V2VRU propagation channel, the LSPs are estimated
in the power and delay domain, i.e., shadow fading (SF), root-mean-square (RMS)
delay spread (DS), and narrowband K-factor (KF), and in the angular domain,
i.e., azimuth spread of departure (ASD), and azimuth spread of arrival (ASA). In
order to maintain the spatial correlation of the LSPs observed in the measured
channel, the autocorrelations of these LSPs are analyzed and the correlation distances
are calculated. Furthermore, to ensure spatial consistency, the cross-correlation
coefficients among the LSPs are calculated. The model parameters are then used as
an input to the simulator.
Chapter 6 presents the validation of the SISO channel model. The simulated
channels are generated by the WINNER-type QuaDRiGa simulator described in
Chapter 2. By comparing the distributions of the model parameters extracted from
the simulated channels with their counterparts extracted from the measured channels
and used as input to the proposed model, the simulated and measured channels are
compared qualitatively. In addition to the distributions of the LSPs, the correlation
distance of each LSP as well as the cross-correlation between each pair of LSPs
is also considered in the validation process. The channel validation reveals that
the proposed model represents the V2VRU propagation channel in the considered
scenarios very well.
Finally, Chapter 7 gives a brief summary of this thesis, and presents future
research directions.

8
 Wave Propagation and Channel
2
Modeling Fundamentals

This chapter provides an overview of wave propagation and channel modeling

fundamentals. The main electromagnetic wave propagation mechanisms are
described. Next, different approaches for channel modeling are briefly discussed and
the recent development in V2VRUs channels are summarized. Finally, a description
of the proposed channel model is provided, followed by a detailed explanation of the
channel coefficient generation procedure.

2.1 Wave Propagation Mechanisms

In wireless communication systems, the emitted electromagnetic wave from the
transmitting antenna travels to the receiving antenna on different paths. This
effect is called multipath propagation and occurs due to different wave propagation
mechanisms such as reflection, scattering, and diffraction. The received signal is
therefore a superposition of a number of attenuated, delayed, and phase-shifted
copies of the transmitted signal. Generally, in real urban environments there are
various objects with different sizes and materials such as trees, buildings, cars,

9
2. Wave Propagation and Channel Modeling Fundamentals

and traffic signs. When an electromagnetic wave travels in an urban environment,

it interacts with these objects and different propagation mechanisms can occur
depending on the sizes and electromagnetic properties of these objects. In the
following paragraphs, these propagation mechanisms are briefly discussed.

Reflection
Specular reflection occurs when the wave impinges on a smooth surface that has
large dimension compared to the incident wavelength [26]. In urban vehicular
communication, reflection usually occurs on the earth surface, buildings, parked
and moving vehicles, and traffic signs. The angle of reflection is equal to the angle
of incidence and the amount of the reflected energy depends on electromagnetic
properties of the material [27, 28], the incident angle, and the wave polarization.

Scattering
Scattering occurs when the electromagnetic wave has a wavelength much larger
than the dimensions of the interacting object in the propagation environment [26].
Scattering may also occur on non-uniform or rough surfaces and in this case it is
called diffuse scattering [29]. Contrary to reflection surfaces in indoor environments,
reflection surfaces in outdoor environments are typically rough. The roughness of
the surface causes the impinging wave to be scattered in non-specular directions.
Furthermore, the non-specular components could be stronger than the specular
component in some cases.

Diffraction
Diffraction describes the bending of waves around edges or corners of objects with
dimensions larger than the wavelength [27]. The diffraction phenomenon can be
explained by Huygen’s principle. According to Huygen’s principle, all points on a
wavefront act as point sources of secondary wavelets which contribute to generate
a new wavefront that propagates into the shadowed region. The power of the
diffracted wave experiences considerable loss. However, the diffracted waves can

10
 2.2 Wireless Channel Models

have a significant contribution to the total received power. The diffraction loss can
be predicted by the single and multiple knife-edge diffraction models [30]. Several
methods are proposed in literature to calculate the diffraction loss based on the
multiple knife-edge diffraction model such as Bullington, Japanese, Epstein-Peterson,
and Daygout methods [31].

2.2 Wireless Channel Models

A channel model aims to represent the wireless propagation channel. It is a tool
used to design, simulate, and analyze a communication system. Through channel
modeling, an understanding of the channel evolution and its properties is acquired
which allows to design a reliable communication system. Channel models can be
broadly divided in three categories: deterministic, stochastic, and Geometry-based
stochastic channel models (GSCMs). In the following subsections, these types of
channel models are briefly described.

2.2.1 Deterministic Channel Models

A deterministic channel model is a site-specific model used to characterize the
propagation channel in a specific environment through simulations. Ray-tracing
(RT) is considered the most popular approach in the deterministic channel modeling
in both indoor and outdoor environments [32, 33]. The RT approach simulates the
reflection and diffraction of waves based on the geometrical optics (GO) model [34].
In GO, the wavelength is assumed to be relatively small when compared with the
dimensions of objects in the propagation environment. The RT is initialized with
the position of the Tx and the Rx, then each path is described by straight lines,
or rays, where all propagation paths between the Tx and the Rx are determined
by geometric considerations. At the Rx side, the received MPCs are characterized
by their amplitude, delay and phase, and then the channel is obtained by the
superposition of all MPCs.
A deterministic approach can produce highly accurate and realistic represen-
tations of the wireless propagation channel in a certain scenario [35]. However, it

11
2. Wave Propagation and Channel Modeling Fundamentals

requires detailed geometrical representation to characterize the channel environment.

Therefore, using RT requires intensive and time consuming computations, and it is
not easily used for system-level simulations. Also, it is very cumbersome to capture
and enter all the relevant details of the physical environment into the ray-tracer,
especially in outdoor environments.

2.2.2 Stochastic Channel Models

Stochastic channel models statistically describe the radio propagation channel by
means of probability distributions derived from the measurement data. Stochastic
channel models, e.g., the tapped-delay line (TDL) model, are widely used for cellular
communications and have been used as a reference model for IEEE 802.11p standard
[36, 37]. The TDL models provide statistics for the received power at a certain
Doppler, delay and angle under the assumption of WSSUS. However, under the
WSSUS assumption, the statistics of each delay tap are time-invariant (WSS) and
are independent between each delay tap (US). Therefore, the TDL models are only
valid within the stationarity region as they lack the ability to describe the transition
between stationarity regions. Therefore, stochastic TDL models are inadequate
to produce realistic models for non-stationary channels such as vehicular channels,
which results in a poor estimation of the communication system performance [13].
The key advantage of stochastic models is that they are mathematically simple and
have relatively low complexity, which makes them easy to use.

2.2.3 Geometry-based Stochastic Channel Models

By combining the geometrical and statistical elements of the deterministic and
stochastic channel models, GSCMs provide accurate geometrical relations at a
reduced computational cost. Moreover, GSCMs are found to be well suited for
non-stationary environments [12]. GSCMs can be further divided into two types,
namely, the European cooperation in science and technology (COST) and the wireless
world initiative for new radio (WINNER). The COST-type GSCMs, on the one
hand, are cluster-centric models that geometrically place random scattering clusters,
drawn from specific distributions, in the simulated propagation environment. These

12
 2.3 State of the Art

distributions are derived from the channel measurement data [14]. These scatterers,
i.e., clusters, fade in and out based on their visibility regions, which in turn depend
on the positions of the Tx and the Rx. The LSPs are then synthesized based on the
visible scatterers to the Tx and Rx. The COST channel models support continuous
time evaluation of the channel. However, the parameterization of the clusters is
challenging. The WINNER-type GSCMs, on the other hand, are user-centric, i.e.,
the placement of the scatterers in the simulated propagation environment is based
on the estimated LSPs. Hence, the LSPs control the behavior of the channel and
describe the distribution of the received power over the delay and angular domains.
Several extensions of the WINNER-type GSCM have been proposed during the
last decade, such as WINNER II [15], WINNER+ [38], 3GPP-3D [39], and the
QuaDRiGa [40] models.

2.3 State of the Art

Several channel models were proposed for V2V and V2I communications [9], [11],
[41], [42]. However, to the best of the author’s knowledge, a dedicated channel
model for V2VRU communication in realistic and critical accident scenarios in urban
environment is still missing.
There are only a handful of recent studies focused on modeling different aspects
of the V2VRU channel. The authors in [43] performed V2P channel measurements
at 3.8 GHz and with a bandwidth of 200 MHz. Based on the measurements in the
LoS scenario, the authors presented a two-ray path loss model for the strongest
path contribution, and log-distance path loss model for the path loss of discrete
scatterers. Normal and Ricean distributions for large and small scale fading were
also reported. The authors extended this work to a LoS/Non-LoS (NLoS) scenario
in [44], where a multi-slope log-distance path loss model is proposed. The authors in
[45] investigated the autocorrelation function of the V2P channel in LoS and partial
LoS scenarios. According to their findings, the channel decorrelates rapidly in time
domain which indicates a non-stationary behavior of the V2P channel. In [46], the
authors investigated the first-order characteristics of the V2P channel at 5.8 GHz in
a business district environment. The K-factor was calculated for different locations

13
2. Wave Propagation and Channel Modeling Fundamentals

of the Tx and Rx antennas and vehicle speeds. However, their work was based on
narrowband received signal strength indication (RSSI) measurements. The path loss
was studied in [47] based on narrowband channel measurements in LoS when the
pedestrian is standing as well as when the pedestrian is moving along the road.
Further studies related to V2VRU communication can be found in the literature.
However, their focus is on evaluating the performance of a V2P communication
system based on narrowband measurements and simulations. For instance, the
authors in [48] conducted a study on the applicability of WiFi-based communication
for V2P scenarios. The authors evaluated the performance of the communication
in terms of packet delivery ratio (PDR) and packet inter-reception time. Based on
experiments, they found out that, in order to satisfy a collision avoidance application,
a transmission rate greater than 1 Hz is required. In [49], the performance of IEEE
802.11p-based V2P communication was evaluated and compared with WiFi and
cellular-based communications. The PDR and end-to-end latency in LoS and NLoS
scenarios were calculated. The authors reported that LTE-based communication
yields better PDR than IEEE 802.11p while the latter one yields a lower latency.
WiFi-based communication shows worse performance than the IEEE 802.11p in
terms of both, PDR and latency. Honda and Qualcomm developed an IEEE 802.11-
based pedestrian safety system in [7]. They implemented a DSRC stack within the
WiFi chipset on a smartphone. To lower the smartphone’s power consumption, a
false-alarm suppression algorithm was developed. The communication performance
was studied in terms of RSSI and inter-reception time. In [8], the use of IEEE
802.11p-based communication between VRUs and vehicles was addressed. The
authors conducted experiments using Cohda MK4 communication units. They found
out that the obstacles located between the Tx and the Rx had a severe impact on
the achievable communication range. The authors in [50] considered an intersection
scenario at which they evaluated the performance of the IEEE 802.11p-based V2P
communication for crash avoidance application through simulations. Their results
showed that even in scenarios with relatively low channel load, there was a significant
loss of packets. According to their results, lower packet inter-reception time can
be achieved by choosing higher-order modulations in IEEE 802.11p. A number of

14
 2.4 Channel Model Description

other studies proposed using cellular-based communication instead of direct ad-hoc

V2VRU communication to enable VRU protection systems (e.g., [51] and [52]).

2.4 Channel Model Description

The modeling approach in this thesis follows the QuaDRiGa modeling approach [40].
the QuaDRiGa model extends the well known WINNER II [15] and WINNER+
[38] models by adding time evolution and 3D propagation. In this model, the
trajectories of both the Tx and the Rx are divided into segments. A channel segment
can be seen as a part of the Tx and the Rx trajectories, in which the LSPs do
not change considerably. The positions of the scatterers are calculated based on
the LSPs. Within a channel segment, the scatterer positions are fixed. However,
due to the movement of the Tx and the Rx within the segment, the propagation
path parameters, i.e, the power, delay, phase, and angles will change. The time
evolution of the channel due to the movement of the Tx and Rx is accomplished
by the so-called drifting and the birth/death of the scatterers. Drifting deals with
the time evolution inside a channel segment. It was introduced in spatial channel
model-extended (SCM-E) [53]. When the Tx and the Rx move from one channel
segment to another, different scatterers will be created. The birth and death process
of the scatterers ensures a smooth transition between adjacent channel segments.

2.4.1 Model Parameters

The model parameters are divided into two sets. The first set of parameters are the
LSPs. The second set contains the supporting parameters.

Large scale parameters : The following five LSPs describe the distribution of
the power over the delay and angular dimensions and control the evolution of the
channel model.

• Shadow fading (SF)

• RMS delay spread (DS)

15
2. Wave Propagation and Channel Modeling Fundamentals

• Ricean K-factor (KF)

• Azimuth spread of departure (ASD)

• Azimuth spread of arrival (ASA)

The distributions of the LSPs are obtained from the parameters of the SMCs,
i.e.,the power, delay, and angles, that are estimated in Sections. 4.5 and 4.6. All
LSPs are modeled with log-normal distribution with mean µ and standard deviation
σ.

Supporting parameters : The following supporting parameters are estimated

from the LSPs and the measured channels.

• Correlation distance of the LSPs

• Cross-correlation of the LSPs

• Number of scattering clusters

• Scaling coefficient for delay distribution

• Scaling coefficient for angle of departure distribution

• Scaling coefficient for angle of arrival distribution

The correlation distance, also called the decorrelation distance, is a scenario-

dependent parameter that determines how long the channel is assumed stationary
for a specific LSP [15, 40]. The cross-correlations between each pair of the LSPs
form a positive definite correlation matrix. They introduce the inter-dependency
between the LSPs. The scaling coefficients are also called the proportionality factors
[15]. They are used to scale path delays and angles to ensure that the differences in
the spreads are reflected in the powers.
All these parameters will be used to generate the channel coefficients that will
be explained in the following subsection. The method used to obtain the LSPs and
the supporting parameters will be addressed in Chapter 5.

16
 2.4 Channel Model Description

2.4.2 Generation of the Channel Coefficients with

QuaDRiGa
The block diagram in Figure 2.1 shows the procedure of the channel generation. In
the first step, the network layout is created by defining the number of the Tx and the
Rx and antenna configurations. The user then needs to provide the trajectories of
the Tx and the Rx. Both the Tx and the Rx move along their trajectories and each
of them has a list of positions defining the trajectory. Since the Tx and the Rx could
move with different speeds, their trajectories may have different lengths. However,
both trajectories have the same number of snapshot positions. The trajectories are
divided into 8 segments with a minimum length of 5 m, an average length of 15 m,
and a standard deviation of 5 m. Since a segment can be seen as an interval in which
the LSPs do not change considerably, these parameters that define the length of
the segments are roughly drawn from the correlation distances of the LSPs that are
evaluated in Chapter 5. The segments of the Tx and the Rx are identical. As an
example, if a segment that has a length of 2000 measurement snapshots begins at
snapshot number 501 and ends at snapshot number 2500 of the Tx trajectory, it
would also start at snapshot number 501 and end at snapshot number 2500 of the
Rx trajectory. The segments are then classified as LoS, obstructed-LoS (OLoS), or
NLoS based on the propagation situation of the channel within the segment in the
considered scenario. For example, in Scenario 1 (see Figure 1.1), the LoS starts to
appear at Tx-Rx distance of 15 m after being obstructed by buildings, therefore, all
segments that correspond to Tx-Rx distance large than 15 m are classified as NLoS,
and the rest are classified as LoS segments.
In the following, the main steps of the channel coefficients generation are
summarized. A more detailed description is found in [15, 40].

A. Calculation of correlated LSP maps

For each LSP, a 2D map is generated. The map gives local values of the LSP at
each Tx and Rx position. Moreover, the spatial correlation of the LSP as well
as its cross-correlation with other LSPs are applied in the map. To clarify, let
us consider the example of the SF discussed in Section 4.2. When the receiver
moves to a shadowed area, it remains shadowed for some time or corresponding

17
2. Wave Propagation and Channel Modeling Fundamentals

Figure 2.1: Steps for calculation of time-evolving channel coefficients.

Figure 2.2: Principle of the generation of correlated LSPs maps.

traveled distance. This implies that shadowing is spatially correlated. The

spatial correlation can be captured by using 2D SF map. As illustrated in
Figure 2.2, based on the Gaussian model, a random (uncorrelated) SF map is
generated (Figure 2.3a). A filter is then applied on the uncorrelated SF map
along x and y axes, and along the two diagonal directions. The used filter is
a decaying exponential filter with coefficients determined from the measured
correlation distance [54]. Compared to the uncorrelated SF map, the resulting
correlated SF map in Figure 2.3b is less fluctuating.

18
 2.4 Channel Model Description

100
15

10
50

Shadow fading [dB]

5
y [m]

0 0

-5
-50
-10

-100 -15
-100 -50 0 50 100
x [m]
(a) Uncorrelated (b) Spatially correlated

Figure 2.3: Examples of uncorrelated and spatially correlated Gaussian shadow

fading maps.

The cross-correlations are applied to the 2D maps by linear transformation

using the measurement-based cross-correlation matrix. Note that, matrix
square root is used instead of Cholesky decomposition for inter-parameter
correlation. The map is then scaled by the mean and standard deviation value
of the LSP. Finally, at a specific location of the Tx and the Rx, the value of
LSP is extracted from the map.
In the next steps B, C, and D, the path delays, powers, and angles are only
calculated for the initial positions, i.e., the beginning of each channel segment.

B. Calculation of initial path delays and angles

The initial delay values for the NLoS paths are randomly drawn from a single-
sided exponential distribution with unit mean and unit standard deviation as
in [40]
τ̃ l = −ln (Xlτ ) , (2.1)

where l is the path index and Xlτ ∼ U (0, 1) is uniformly distributed and
spatially correlated random variable that has values between 0 and 1. In this
step, the LoS path is assigned with a zero delay.

19
2. Wave Propagation and Channel Modeling Fundamentals

In the next step, the two azimuth angles are initialized for each NLoS path.
The angles are drawn randomly from a uniform distribution as

π π
(︃ )︃
ϕ̃l = Xlϕ
∼U − , , (2.2)
2 2
while the initial angles for the LoS path are set to 0. Under the assumption
that all scatterers are located on the same horizontal plane with the Tx and
the Rx antennas, the elevation angles of the paths have zero values.

C. Calculation of initial path powers

d
Based on the initial delays τ̃ l , and initial azimuth angles of departure ϕ̃l and
a
arrival ϕ̃l , the initial path powers are calculated as [40]

P˜l = exp{−τ̃ l · g DS − (ϕ̃l )2 · g ASD − (ϕ̃l )2 · g ASA },

d a
(2.3)

where g DS ,g ASD , and g ASA are the scaling coefficients used to ensure that the
differences in the spreads are reflected in the powers. Differently from [40], the
scaling factors are calculated based on their definitions in [15]. The scaling
factor of the delay is defined as the ratio between the standard deviation of
the path delays and the RMS delay spread. Similarly, the scaling factor of
the angles is defined as the ratio between the standard deviation of the path
angles and the angle spread.

D. Applying K-factor, delay spread and angle spreads

After initializing the path delays, angles, and powers, the actual values of
the K-factor, delay spread and the two angular spreads (AS) are applied. By
definition, the K-factor K is the ratio of the power of the LoS path to the sum
of the power of all other paths. Therefore, the power of the LoS path is scaled
as

L
P̃ 1 = K · (2.4)
∑︂
P̃ l ,
l=2

20
 2.4 Channel Model Description

with L being the number of total paths. The path powers are then normalized
to have a sum power equals one Watt as

L
Pl = P̃ l / (2.5)
∑︂
P̃ l ,
l=1

After that, the scaled path powers Pl from Equation (2.5) and the initial path
delays τ̃ l from Equation (2.1) are used to calculate the initial delay spread as
⌜
⃓ )︄2
⃓1 L
1 ∑︂
L
⃓ (︄
σ˜τ = Pl · (τ̃ l )2 − (2.6)
∑︂
⎷ · · Pl · τ̃ l ,
P l=1 P l=1

where P is the sum of all path powers. Next, by applying the initial delay
spread σ˜τ and the real value of the delay spread στ from the correlated 2D
map, the initial path delays from Equation (2.1) are scaled to obtain the a
new path delays as

στ
τl = τ̃ l · . (2.7)
σ˜τ

By using these path delays together with the path powers from Equation (2.5)
the correct delay spread can be achieved.
The angular spread is a measure of the spread of the path powers in angular
domain. Similar to applying the delay spread, the initial angular spread is
calculated by
⌜
⃓ )︄2
⃓1 L
1 ∑︂
L
⃓ (︄
(︂ )︂2
σ˜ϕ = (2.8)
∑︂
⎷ · Pl · ϕ̃l − · Pl · ϕ̃l ,
P l=1 P l=1

then the initial angles ϕ̃l from Equation (2.2) are scaled using the correct
angular spread σϕ from the 2D correlated map as
(︂ )︂ σϕ
ϕl = exp j · ϕ̃l · s , s= , (2.9)
σ˜ϕ

21
2. Wave Propagation and Channel Modeling Fundamentals

where the exp function is used to warp the angles around the unit circle.
As mentioned earlier, the azimuth angles of departure ϕd1 and arrival ϕa1 of
the LoS path were set to 0. The correct angles are now calculated from the
positions of the Tx and the Rx as

ϕd1 = arctan2 {yr − yt , xr − xt } , (2.10)

ϕa1 = ϕdl + π . (2.11)

Next, the path angles are converted from spherical to Cartesian coordinates by
⎛ ⎞
cos ϕl ⎠
cl = ⎝ . (2.12)
sin ϕl

Then, the angles are rotated around the y-axis using a rotation matrix
constructed from the LoS angles as
⎛ ⎞
cos ϕ1 − sin ϕ1 ⎠
ĉl = ⎝ · cl . (2.13)
sin ϕ1 cos ϕ1

In the last step, the final angles are obtained by converting ĉl back to spherical
coordinates as
ϕl = arctan2 {ĉl,x , ĉl,y } , (2.14)

where arctan2 is the multi-valued inverse tangent. Note that, in the previous
steps the elevation angles are ignored as they are set to 0.

E. Drifting of delays, angles and phases over each segment

Until now, the path delays, powers, and angles are calculated at the initial
positions of the Tx and the Rx, i.e., at the beginning of each channel segment.
When the Tx and the Rx move, the values of the path delays, power, and
angles should be updated accordingly.

22
 2.4 Channel Model Description

Figure 2.4: Illustration of the calculation of the scatterer positions and updates of
the departure and arrival angles in the multi-bounce model.

A double-bounce model, illustrated in Figure 2.4, is considered in implementing

the channel model. Therefore, the NLoS path consists of three parts. In the
first part, the vector bl points from the Tx antenna to the first-bounce scatterer
(FBS). In the second part, the vector cl points from the FBS to the last-bounce
scatterer (LBS), and in the third part, the vector al points from the Rx antenna
to the LBS. The total path length dl is calculated from the path delay τl at
the initial position, i.e., at the beginning of the segment. Hence, the total path
length is
dl = |bl | + |cl | + |al | . (2.15)

Calculating the position of the FBS and the LBS requires the lengths of the
vectors al and bl . One way to calculate |al | and |bl | is by solving the following
optimization problem with an objective to minimize the length |cl | as follows

minimize |cl | = dl − |bl | − |al |

|al |,|bl |

subject to r = b̂l · |bl | + ĉl · |cl | − âl · |al | ,

(2.16)
|bl | ≥ dmin ,
|al | ≥ dmin ,

23
2. Wave Propagation and Channel Modeling Fundamentals

where dmin is the minimum distance between the antenna and the nearest
scatterers, and it is introduced in order to obtain realistic results. The unit-
length vectors b̂l and âl are calculated from the departure and arrival angles.
The ideal solution of the optimization problem is obtained when |cl | becomes
zero and then the double-bounce model turns into a single-bounce model.
During a segment, the positions of the scatterers stay fixed and used to update
the path delays and angles at each measurement snapshot.
At snapshot s, the arrival angle is obtained by converting back to spherical
coordinates as
ϕal,s = arctan2 {al,s,y , al,s,x } , (2.17)

where al,s is a vector pointing from the Rx location at snapshot s to the LBS.
Similarly, the departure angle is calculated at snapshot s by

ϕdl,s = arctan2 {bl,s,y , bl,s,x } , (2.18)

where bl,s is a vector pointing from the Tx location at snapshot s to the FBS.
The path phases ψl,s and delays τl,s are calculated from the total path length
at snapshot s as follows

dl,s = |bl,s | + |cl,s | + |al,s | , (2.19)

2π
ψl,s = · (dl,s mod λ) . (2.20)
λ

dl,s
τl,s = , (2.21)
c
with c being the speed of light. Finally, the LoS angles are updated as

ϕd1,s = arctan2 {rs,y , rs,x } , (2.22)

24
 2.4 Channel Model Description

ϕa1,s = arctan2 {−rs,y , −rs,x } , (2.23)

where rs is a vector pointing from the Tx location to the Rx location at

snapshot s.

F. Application of path gain, shadow fading and K-factor

The path gain PGs at snapshot s is derived from the path loss models
introduced in Section 4.1. The SF is obtained from the 2D correlated map.
The effective path gain is obtained by combining the PG and the SF as
√︃ (︂ )︂
[dB] [dB]
0.1 PGs +SFs
PG[eff]
s = 10 . (2.24)

As the KF is spatially correlated, it changes its value when the Tx and the
Rx change their position. Previously, the path powers at the beginning of the
segment Pl are scaled by the initial KF in Equations (2.4) and (2.5). However,
since the KF changes, additional scaling factor for the path powers is required
⎧√︂
⎪ Ks
for l = 1;
√︄
Ks
(︃ )︃ ⎨
[scale]
Kl,s = 1 + P1 −1 ·⎪ K0
(2.25)
K0 ⎩1 otherwise ,

where K0 is the KF at the beginning of the segment, Ks is the KF at snapshot

s obtained from the 2D correlated map, and P1 is the power of the LoS path
calculated in Equation (2.4).
Finally, the complex-valued channel coefficients are calculated by
√︂
[scale]
gl,s = PG[eff]
s · Kl,s · Pl · e−jψl,s . (2.26)

G. Transition between segments

In the previous steps, scatterrers, channel coefficients, and path delays are
created independently for each segment. Within each segment the scatterring
clusters remain fixed, and the LSPs change slowly. When the Tx and the Rx
move from one segment to another, different scatterers will be created, and

25
2. Wave Propagation and Channel Modeling Fundamentals

the LSPs will change considerably. However, in reality, the physical channel
does not change rapidly when moving between adjacent segments. In order to
allow the Tx and the Rx move longer distances and maintain LSPs consistent,
longer time evolution of the channel is required. The time evolution is achieved
by the birth and death process of the scatterers. This process is modeled by
merging the scatterers, i.e., the NLoS paths of adjacent segments.

Figure 2.5: Illustration of the overlapping area used for calculating the transitions
between segments.

As depicted in Figure 2.5, segment 2 is extended to overlap with segment

1. Within the overlap gray area, the extension of each scattering cluster,
i.e., NLoS path is depicted with a dashed blue line. Therefore, the resulting
scatterring clusters have lifetimes restricted to the combined length of two
adjacent segments. The overlapping area is split into multiple parts equal
to the number of the scatterers, e.g., three parts in this example. Within
each overlapping part, one old NLoS path is paired with one new path. The
selections of the two paths to be merged is done carefully based on their powers
and delays. The selection criteria is minimizing the sudden fluctuation of the
LSPs when moving between the two segments. Then, the power of the old
path ramps down and the power of the new path ramps up. The power ramps
are modeled by a squared sine function [40]

26
 2.4 Channel Model Description

π
(︃ )︃
w[sin] = sin2 · w[lin] , (2.27)
2
where w[lin] is the linear ramp with range from 0 to 1, and w[sin] is the sine-
shaped ramp.

Finally, the proposed channel model is validated where the simulated channels,
generated by the previous steps, and the measured channels are compared
qualitatively in Chapter 6.
However, in order to estimate the model parameters and generate the channel
coefficients, channel measurements need to be conducted. Extensive wideband
channel measurements were performed in both open-field and urban environments.
A detailed description of the two channel measurement campaigns is presented in
the next chapter.

27
 Wideband Channel Measurements
3
In order to obtain the parameters required for modeling the V2VRU communication
channel, measurement data needs to be recorded. In this chapter, two measurement
campaigns are described. The first campaign was executed in an open-field
environment and presented in Section 3.1. The data collected in this campaign is
later used to study the impact of several aspects on the propagation channel, e.g.,
Tx and Rx mobility, and LoS blockage by parked vehicles. In Section 3.2, the second
measurement campaign in an urban environment considering the three most critical
accident scenarios is described. The proposed channel model in this thesis is based
on data collected during the measurements in the urban environment.

3.1 Open-field Measurements

To get detailed insight into the propagation of electromagnetic waves between vehicles
and VRUs, we performed a measurement campaign in March 2017 at the airport
in Oberpfaffenhofen near Munich. This location was chosen since it represents a
controlled environment with open-sky visibility and a small number of far-located
objects that could potentially reflect or scatter the electromagnetic waves. This
location makes it possible to isolate and study the impact on the propagation channel

29
3. Wideband Channel Measurements

caused by the different elements in the propagation environment as well as by the

mobility of the Tx and Rx. For our measurement campaign, a collision scenario
of a vehicle and a pedestrian was considered. Here, the test vehicle drove straight
towards the pedestrian, which was approaching an imaginary collision point, from
the right side. Both the pedestrian and the vehicle met at an imaginary collision
point. Figure 3.1 displays the trajectory of both the test vehicle and the pedestrian
on an aerial view of Google Maps. In our experiments, the test vehicle acted as the
transmitter, while the pedestrian played the receiver role. The next two sections
describe the measurement systems and scenarios.

Figure 3.1: Aerial view of the measurement scenario showing the trajectories of the
TX and the RX towards the imaginary collision point. (Google Maps 2017 Geobasis-
de/BKG.)

3.1.1 Measurement Systems

The wideband channel sounding measurements were performed using the RUSK-
DLR channel sounder at a center frequency fc = 5.2 GHz. The measurements
bandwidth was B = 120 MHz, which corresponds to a delay resolution of ∆τ =
8.33 ns. The time-variant channel transfer function was recorded every Tg = 1.024 ms,
which allows to record a maximum absolute Doppler frequency of fd = 488 Hz. To
fulfill the requirement of maximum Doppler frequency, the vehicle velocity was
restricted to vmax = 11 m/s. The length of each time-variant channel transfer
function snapshot was Tp = 0.8 µs. The configuration parameters of the channel

30
 3.1 Open-field Measurements

Parameter Value
RF center frequency fc 5.2 GHz
Bandwidth B 120 MHz
Transmit power 37 dBm
Signal period Tp 0.8 µs
Time grid Tg 1.024 ms
Tx antenna Onmi-directional (V-polarized), 8 dBi
Rx antenna Omni-directional (V-polarized), 8 dBi and
Dual-polarized array
Vehicle speed 11 m/s
Pedestrian speed 1.2 m/s

Table 3.1: Channel sounder parameters in the open-field measurements.

sounder are summarized in Table 3.1. The transmit antenna was positioned at the
front side of the roof of the vehicle at a height of hTx = 1.9 m above the ground. The
receive antenna was placed either on a tripod or carried by a pedestrian at heights
of hRx = 1.1 m or 1.3 m.
In order to synchronize the transmitter and the receiver, two rubidium clocks
were used. However, during post processing, it was found that a clock offset had
been accumulated during the time of the experiments. In order to compensate this
offset drift, the relative position of the Tx and Rx was used to calculate the difference
between the propagation delay of the LoS path in the CIR and the true propagation
delay. The drift values were calculated in all experiments. It was found that the offset
increased linearly during the measurements day. Based on the linear increase, the
value of the offset for each measurement snapshot was calculated and compensated.
GNSS was used as a ground truth for the position of both transmit and receive
antennas. The vehicle, the tripod and the pedestrian were equipped with a Topcon
Legacy E+ L1/L2 GLONASS/GPS receiver. A geodetic-grade GNSS antenna was
placed on the roof of the vehicle, on one end of the tripod and attached to the helmet

31
3. Wideband Channel Measurements

of the pedestrian (see Figure 3.2). The 10 Hz recorded GPS and GLONASS raw data
were post processed to find a carrier-fixed solution with centimeter-level accuracy
(1σ). The displacement between GNSS and the communication antennas at Rx and
Tx was considered when computing the exact position for the propagation analysis
with the channel sounder and the ITS-G5 system. To determine the location of
the parked vehicles acting as obstruction to the LoS, an LD-MRS multi-layer laser
scanner from Sick was attached at the front bumper of the test vehicle and a ublox
LEA 4T GPS receiver was employed. In post-processing, the laser point cloud was
transformed from a vehicle coordinate frame to a global coordinate frame using the
code-solution from the GPS receiver.

3.1.2 Measurement Scenarios

The test vehicle in which the transmitter was located was a Mercedes G400
(Figure 3.2a). After an initial acceleration phase, the vehicle moved from 100 m
distance towards the collision point with constant speed of 11 m/s . To study the
influence of movement and body shadowing of the pedestrian, we performed tests
with a static and with a moving pedestrian. In the static case, the pedestrian
(Figure 3.2b) was replaced by a tripod (Figure 3.2c). The tripod was placed at three
different positions with different distances (12 m, 7 m and 2 m) from the collision
point. In the moving case, the pedestrian was approaching the collision point from
12 m distance at a speed of approximately 1.2 m/s.

32
 3.1 Open-field Measurements

(a) Tx Test Vehicle (b) Rx Pedestrian

(c) Rx Tripod (d) Rx Antenna Array

Figure 3.2: Transmitter (Tx) and receiver (Rx) involved in the open-field
measurement campaign.

To study the impact of the parked vehicles at the roadside on the propagation
channel especially the blockage of the LoS path, a row of cars and vans were parked
in a line parallel to the trajectory of the test vehicle. Five different constellations
using one to six vehicles of different size and shape were tested. Additionally, one of
the cars was used to study the effect on the communication channel of an additional
reflection coming from the opposite roadside, once, from a parked car and, once,
from a moving car. Further, to study the effect of the shadowing of a crowd of people
surrounding the pedestrian, the pedestrian was surrounded by four test persons. In
addition, a circular antenna array with 2 × 16 elements was used (See Figure 3.2d)

33
3. Wideband Channel Measurements

in order to detect the angle of arrival of different multipath components and study
scattering phenomena.

(a) Senario 1: LoS with tripod (b) Senario 2: LoS with moving pedestrian

(c) Senario 3: Shadowing crowd (d) Senario 4: NLoS with static tripod

Figure 3.3: Four measurement scenarios in the open-field environment addressed in

this thesis.

All in all, a total set of 30 experiments with different combinations of

moving/static pedestrian, LoS/NLoS, different combinations of parked vehicles
and with/without crowd shadowing were performed. In this thesis, the following
four scenarios will be discussed in details:

• Scenario 1 (Figure 3.3a): LoS condition with static tripod

• Scenario 2 (Figure 3.3b): LoS condition with moving pedestrian

• Scenario 3 (Figure 3.3c): LoS condition with shadowing crowd

34
 3.2 Urban Measurements

• Scenario 4 (Figure 3.3d): NLoS condition with static tripod

3.2 Urban Measurements

Channel measurements were conducted in October 2018 at three different streets
in the city of Germering near Munich. In this thesis, the channel measurements in
Göthestraße will be analyzed (see Figure 3.4) where 3 −6 story buildings lined up
along the street on one side, while separated by green area on the other side. The
street consists of one lane for each direction with parked cars on both sides. It is
12 m wide, with 3 m wide sidewalks.

Figure 3.4: Aerial view of the measurement scenarios in the urban environment show-
ing the trajectory of the vehicle (in blue), and the trajectory of the cyclist/pedestrian
(in green with the number of the scenario) towards the imaginary collision point.
(Google Earth 2018 Geobasis-DE/BKG.).

35
3. Wideband Channel Measurements

3.2.1 Measurement Systems

The V2VRU measurement campaign was carried out using the RUSK-DLR channel
sounder. The measurement signals were transmitted at a carrier frequency of fc =
5.2 GHz, which is close to the 5.9 GHz ITS band. The bandwidth was B = 120 MHz,
and thus providing a delay resolution of ∆τ = 8.33 ns. During the measurements,
the time-variant channel transfer function was recorded every Tg = 1.024 ms allowing
to resolve a maximum Doppler shift of fd max = 488 Hz. The main measurement
parameters can be found in Table 3.2. In order to record the position of the Tx and
the Rx antennas, GNSS receivers were used. Two lidar sensors from Velodyne were
mounted on the test vehicle and near the cyclist/pedestrian to provide 3-dimensional
high resolution images of the environment. The antennas positions on Tx and Rx
side can be seen in Figure 3.5.

Figure 3.5: Antennas positions on the pedestrian, the cyclist, and the test vehicle.

3.2.2 Measurement Scenarios

Figure 3.4 displays the trajectories of the test vehicle, the cyclist, and the pedestrian
on an aerial view for the accidents scenarios. Three critical accident scenarios
between vehicle and VRU were considered:

• Scenario 1 (Figure 3.6a): The vehicle is moving 100 m with an average velocity
of 5 −11 m/s towards the intersection then turning right while the cyclist is
moving 10 m with an average speed of 1.5 m/s toward the imaginary collision

36
 3.2 Urban Measurements

Parameter Value
RF center frequency fc 5.2 GHz
Bandwidth B 120 MHz
Transmit power PT 37 dBm
Signal period Tp 3.2 µs
Time grid Tg 1.024 ms
Tx antenna Onmi-directional (V-polarized), 8 dBi
Rx antenna Omni-directional (V-polarized), 8 dBi
Vehicle speed 5 −11 m/s
Cyclist speed 1.5 m/s
Pedestrian speed 1 m/s

Table 3.2: Channel sounder parameters in the urban measurements.

point from the right. This scenario accounts for 42 % of the total cyclist
accidents as reported by the general association of the German insurance
industry (GDV) [3]. In this scenario, the LoS between the vehicle and the
cyclist is initially blocked by buildings, then it becomes obstructed by parked
vehicles as both the vehicle and the cyclist are approaching the collision point.
In this scenario, a total of 10 measurement runs were conducted.

• Scenario 2 (Figure 3.6b): In contrast to Scenario 1, the cyclist is moving

parallel to the vehicle toward the collision point at the intersection. This
scenario is less critical than Scenario 1 as it accounts for only 11 % of the
total cyclist accidents [3]. In this scenario, the LoS between the vehicle
and the cyclist repeatedly transits between LoS and OLoS situations due to
obstruction by parked vehicles. In this scenario, a total of 2 measurement runs
were conducted.

• Scenario 3 (Figure 3.6c): The pedestrian is crossing the street with an average
speed of 1 m/s while the vehicle is approaching from the left. In this scenario,

37
3. Wideband Channel Measurements

the visibility between the vehicle and the pedestrian is partially or completely
blocked by the parked vehicles. According to the national highway traffic
safety administration (NHTSA) [2], this scenario is the most critical accident
scenario for pedestrians and it accounts for 26 % of all pedestrian accidents.
In this scenario, a total of 9 measurement runs were conducted.

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 3.6: Illustrations of the accident scenarios in urban environment. The data
collected in these scenarios are used to model the V2VRU channel.

38
 Channel Characteristics and General
4
Modeling Aspects

This chapter addresses several aspects of channel modeling. In Section 4.1, path loss
models for open-field and urban channels are proposed. Moreover, due to the spatial
correlation of the shadow fading, consecutive packet losses can occur, which lead to
degradation in the communication performance. Therefore, the spatial correlation of
the shadow fading in the open-field is analyzed and models are proposed. To study
the propagation loss due to the obstruction of the LoS by parked vehicles, a 3D
ray tracing tool has been developed in Section 4.3. The tool detects the diffraction
edges and calculates the Fresnel-Kirchoff parameter that is used to calculate the
knife-edge diffraction loss.
Due to the non-stationarity of the V2VRU channel, the channel is characterized
by dividing it into regions where the WSSUS assumption holds and then the LSPs
are estimated in each individual region. In Section 4.4, the non-stationarity of
the V2VRU channel is analyzed by estimating the GLSF and its collinearity based
on the channel measurement data in the urban environment, and the estimated
stationarity distance is presented.

39
4. Channel Characteristics and General Modeling Aspects

In Section 4.5, the multipath parameters are estimated using the KEST algorithm.
A novel method is proposed to extract SMCs from the estimated CIR based on the
density of their neighboring MPCs. This extraction allows for further characterization
of these SMCs. Furthermore, a simple but effective algorithm for multipath tracking
based on the differences in delay and magnitude between SMCs is presented. Based
on the previous step, in Section 4.6, the SMCs parameters are employed to localize
the physical scatterers in the propagation environment using a joint delay-Doppler
estimation algorithm. The estimated positions of the scatterers are then used to
estimate the AoD and AoA. In Section 4.7, the diffuse multipath is extracted in
order to calculate its contribution to the total received power.

4.1 Path Loss Model

In this section, path loss models for the measured scenarios in the open-field and
urban environments are presented.
The CIR can be expressed mathematically as a sum of N (t) Dirac impulses [55]:

N (t)−1
h(t, τ ) = αi (t) · δi (τ − τi (t)), (4.1)
∑︂

i=0

where αi (t) and τi (t) are the complex amplitude and delay of the i-th MPC at time
step t and δ(·) is the Dirac function. The instantaneous power delay profile (PDP)
is then calculated by:
P (t, τ ) = |h(t, τ )|2 , (4.2)
√︂
where |h(t, τ )| = (Ih (t, τ ))2 + (Qh (t, τ ))2 , where Ih (t, τ ) and Qh (t, τ ) are the real
part and the imaginary part of the complex amplitude αi (t) [56]. Next, the average
PDP (APDP) P (t, τ ) is obtained by applying a sliding window1 on the PDP with
a length corresponding to a traveled distance of lw =10 λ which is equivalent to
Nav = v·T
lw
g
= 54 PDPs in the open-field measurements and 169 PDPs in the urban
measurements, where v =11 m/s is the average velocity of the transmitter in open-
1
Different window sizes are tested, a window size of 10 λ is found to best fit the data. It provides
a sufficient number of samples to accurately extract the small-scale fading.

40
 4.1 Path Loss Model

field measurements and it is equal to 7 m/s in the urban measurements, Tg =1.024 ms

is the time grid, and λ is the wavelength. Within each segment of 10 λ, the channel
is assumed to be quasi-stationary.
Based on the APDP, the received power Pr (t) can be calculated as:

N (t)−1
Pr (t) = P (t, τi ) , (4.3)
∑︂

i=0

where N (t) is the number of the APDP samples at time t. Thereafter, the path loss
PL is obtained by:
Pt GTx GRx
PL = , (4.4)
Pr
where Pt , Pr , GTx , and GRx are the transmitted power, received power, Tx antenna
gain and Rx antenna gain, respectively.
The path loss can be expressed in logarithmic scale as a sum of two components,
namely, the distance-dependent path loss model and the SF:

PL = PL (d) + X
⏞⏟⏟⏞ . (4.5)
⏞ ⏟⏟ ⏞
path loss model shadow fading

The SF is caused by shadowing processes and leads to a change in the local mean of
the path loss over relatively large distances. Whereas the small-scale fading is the
variation due to the superposition of multiple propagation paths. It leads to power
variations when moving over relatively short distances in the order of the signal
wavelength. The small-scale fading is extracted by applying the aforementioned
sliding window.
The log-distance path loss model is used to predict the propagation loss in
different environments. The average large-scale path loss is a function of the
separation distance between the Tx and Rx as expressed by

PL (d) = PL (d0 ) + 10nlog10 (d/d0 ), (4.6)

where PL (d) denotes the path loss in dB at a distance d between Tx and Rx, PL (d0 )
is the path loss at a reference distance d0 , and n is the path loss exponent. Using

41
4. Channel Characteristics and General Modeling Aspects

linear regression analysis, the path loss exponent n that minimizes the difference
between the measured and modeled values is determined.
In scenarios where the LoS becomes partially or completely obstructed, it has
been found that the measured path loss can be more accurately represented by a
dual slope log-distance path loss model. The dual slope model is given by
⎧
⎨PL (d0 ) + 10n1 log
10 (d/d0 ), if d ≤ dc
⎪
PL (d) = (4.7)
⎩PL (d0 ) + 10n1 log
10 (dc /d0 ) + 10n2 log10 (d/dc ), if d > dc .
⎪

where PL (d0 ) is the path loss at the reference distance d0 , and dc is the breakpoint
distance at which the second slope begins. n1 and n2 are the path loss exponents
of the first and second slope, respectively, and they provide an indication of how
the signal attenuates with the Tx-Rx separation distance. The breakpoint distance
corresponds to the Tx-Rx separation distance at which the communication link
state changes, i.e., from LoS to OLoS or from OLoS to NLoS. The model fitting
parameters were tuned by minimizing the squared error between the measured path
loss and the model (i.e., least squares criteria).
Another path loss model, widely used in literature to model the LoS channel,
is the two-ray path loss model [30]. This model considers not only the direct
propagation path but also the reflected one from the ground. The received power
according to the two-ray path loss at a Tx-Rx separation distance d is given by

4π ⃓ e−jkdLoS e−jkdgr ⃓⃓
⃓ ⃓
PL (d) = 20log10 ( ) − 20log10 ⃓⃓ + Γ(θ) ⃓, (4.8)
⃓
λ dLoS dgr ⃓

where the parameter k = 2π/λ √︂

is the wavenumber at the center
√︂
frequency fc and
λ is the wavelength. dLoS = d2 + (hTx − hRx )2 and dgr = d2 + (hTx + hRx )2 are
the propagation distances for the LoS and the ground reflection. hTx and hRx
are the heights of the Tx antenna and the Rx antenna, respectively. Γ(θ) is the
ground reflection coefficient for the reflection angle θ and is calculated for vertical
polarization by √︂
ϵr sin(θ) + ϵr − cos2 (θ)
Γ(θ) = √︂ (4.9)
ϵr sin(θ) − ϵr − cos2 (θ)

42
 4.1 Path Loss Model

where ϵr is the relative permittivity of the ground and θ = tan−1 ( hTx +h

d
Rx
).
Figure 4.1 shows the measured path loss and the proposed path loss models for
the open-field measured scenarios described in Section 3.1.2. Note that the free
space path loss model has n = 2 and PL (d0 ) = 46.77 dB, where d0 = 1, and is plotted
for comparison. Table 4.1 summarizes the results of fitting the measured path loss
linearly to the proposed log-distance models.

(a) Scenario 1: LoS with tripod (b) Scenario 2: LoS with pedestrian

(c) Scenario 3: shadowing crowd (d) Scenario 4: NLoS

Figure 4.1: The measured and modeled path loss in the open-field environment.

In Scenario 1, the receiver antenna is fixed on a static tripod at height hTx = 1.1 m.
The tripod was 7 m away from the collision point (see Figure 3.3a). Figure 4.1a
shows the obtained path loss versus the distance between the Tx and Rx. A typical

43
4. Channel Characteristics and General Modeling Aspects

two-ray effect, i.e., the LoS path and the ground reflected path, can be noticed which
causes a power variation slowly over the distance between the Tx and Rx. With
an estimated relative permittivity ϵr = 1.05 for the ground and standard deviation
σ = 0.68 dB, the two-ray model provides a good fit to the measurement data even
at short distance between the Tx and Rx.
In Scenario 2, the receive antenna was mounted with a moving pedestrian (see
Figure 3.3b). Figure 4.1b shows the corresponding path loss versus the distance
between the Tx and Rx. The two-ray model does not fit to the measurement data
and, therefore, is not visualized. A rapid fluctuation of the measured path loss can
be seen. This fluctuation is probably the result of the MPC originated by the human
body. Further, the moving human body also results in dynamic antenna altitude
and position. The estimated log-distance path loss exponent n is 2.44, which is
larger than the n for the free space propagation.
In Scenario 3 (see Figure 3.3c), the shadowing influence of neighboring pedestrians
on the received power is evaluated. Figure 4.1c shows the obtained path loss from
the measurement. Due to the shadowing effect caused by the crowd around the
Rx, the two-ray effect can not be clearly observed. In general, the measured path
loss value is 5 −10 dB larger than the LoS case. The estimated path loss exponent
n is 1.26, which is smaller than the n in Scenario 1. It indicates that the power
in Scenario 3 decreases slower than in Scenario 1, however, with more shadowing
caused by the pedestrians, i.e., a higher value of PL (d0 ) = 67 dB. A similar finding
is also reported in [57].
In Scenario 4, the receive antenna was stationary next to a convoy consisting of
6 parked vehicles with different sizes. Adjacent parked vehicles are separated with
a gap of 1 m (see Figure 3.3d). During the movement of the Tx toward the Rx,
the LoS is partially or completely obstructed. Figure 4.1d shows the measurement
result for the path loss values. The measurement samples between 60 and 100 m are
obtained under LoS condition without obstruction by the parked vehicles. It can be
seen that the path loss value is similar to the LoS condition in Figure 4.1a between
60 and 100 m.
It can be noticed that compared to the free space path loss model, the obstruction
of the LoS causes an extra loss between 10 − 20 dB depending of the size of the

44
 4.1 Path Loss Model

n PL (d0 ) [dB] σ [dB]

Scenario 2 2.44 40 3.20

Scenario 3 1.26 67 5.47
Scenario 4 - LoS 2 46.77 4.35
Scenario 4 - OLoS 1 73 4.35

Table 4.1: Log-distance path loss model parameters in the open-field environment.

parked vehicle. For Tx-Rx distances between 7 m and 60 m the path loss experiences
large values due to obstruction by the parked vehicles. Therefore, the path loss model
is divided into LoS and OLoS parts. Each part has a different path loss exponent n
and standard deviation σ. The parameters of the log-distance path loss model for
both the LoS scenario and the LoS part of the NLoS scenario are similar to those
of the free space model. On the other hand, the path loss exponent n = 1 during
the obstruction of the LoS indicates that the power decreases slower than during
LoS situation, however, with more shadowing caused by the parked vehicles, i.e., a
higher value of PL (d0 ) = 73 dB. A similar inverse proportional relation between n
and PL (d0 ) is reported in [22]. The diffraction loss due to blockage of the LoS by
the parked vehicles is investigated and modeled in the next section.
Next, path loss models for the measured path loss in the urban environment
are provided. Figure 4.2 shows the measured path loss variations with the Tx-Rx
distance and the proposed path loss models for the measured scenarios in the urban
environment. Table 4.2 summarizes the results of fitting the measured path loss
linearly to the proposed models.
In Scenario 1 (Figure 4.2a), due to movement of the vehicle (Tx) and the cyclist
(Rx), the Tx-Rx link transits between NLoS, OLoS, and LoS. A NLoS situation
arises due the blockage of the LoS by buildings between 100 and 32 m. Due to the
obstruction, an extra loss up to 25 dB is experienced when compared to the free
space path loss. Between 32 m and 15 m the LoS becomes obstructed by the parked
vehicles which causes a diffraction loss of 5 −15 dB. A LoS state appears only shortly
before the collision between the vehicle and the cyclist at the intersection. When

45
4. Channel Characteristics and General Modeling Aspects

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 4.2: The measured and modeled path loss in the urban environment.

46
 4.1 Path Loss Model

compared to the free space path loss, a difference up to 7 dB can be noticed. This
difference is partially due to the ground and self-body reflection. The measured
path loss is found to be more accurately represented by a dual-slope log-distance
path loss model given by Equation (4.7) for distances between 15 and 100 m with a
breaking distance of 55 m. Path loss exponents of 6.9 and 2.9 for the first and the
second slopes, respectively, are proposed. The path loss during LoS is modeled by a
single slope log-distance model with an estimated path loss exponent of 1.4.
In Figure 4.2b, the path loss for Scenario 2 is modeled by two single slopes
log-distance models. The first slope is for distances up to 9 m and covers the LoS
area prior to the collision with estimated path loss exponent of 1.7 . For distances
between 9 m and 100 m, the link between the vehicle and the cyclist becomes
partially obstructed by parked vehicles which causes a diffraction loss up to 5 dB.
The estimated path loss exponent equals 2.4, which is slightly above the one of the
free space model.
In Scenario 3 (Figure 4.2c), the estimated path loss exponent for LoS area, i.e.,
up to 9 m, does not vary much from Scenario 2 with an exponent of 1.8 . However,
the path loss experiences a sudden increase by 15 dB at around Tx-Rx distance of
10 m due to the obstruction of the LoS by a large vehicle. The path loss between
9 m and 100 m shows rapid variations in the order of 5 to 15 dB due to obstruction
of the LoS by parked vehicles with an estimated path loss exponent of 3.2.

LoS OLoS \NLoS

Slope 1 Slope 1 Slope 2
P L(d0 ) [dB] n σ [dB] P L(d0 ) n σ n σ

Scenario 1 55.70 1.40 1.90 -15.16 6.90 3.80 2.90 3.46

Scenario 2 49.46 1.70 1.53 44.37 2.40 1.88
Scenario 3 49 1.80 3.40 31.97 3.20 5.79

Table 4.2: Log-distance path loss model parameters for urban scenarios.

47
4. Channel Characteristics and General Modeling Aspects

4.2 Spatial Correlation of Shadow Fading in Open-

field
The SF is strongly dependent on the environment, i.e., it varies from one area to
another. Movements of the vehicle, the VRU, and the surrounding scatterers result
in a dynamic environment. When the receiver moves to a shadowed area, i.e., the
LoS becomes obstructed, it remains obstructed for some time or corresponding
traveled distance. This implies that the shadowing is spatially correlated. Due to
this underlying correlation, consecutive packet losses can occur during this time
interval. As a result, communication performance will be degraded. Therefore,
accurate modeling of the spatial correlation of the SF is important to design a
reliable V2VRU communication system. Spatial correlation models are used in
simulations to generate realizations of the SF process with the desired correlation.
Different techniques for generating the SF based on the correlation models can be
found in literature, such as [58] and [59].
The SF is extracted by subtracting the path loss model from the measured path
loss after removing the impact of the small-scale fading by averaging the measured
path loss using a sliding window of length lw =10 λ. This window size is found to
provide a sufficient number of samples to accurately extract the small-scale fading
and calculate the correlation distance of the SF. Figure 4.3 shows the extraction of
the SF from the measured path loss for Scenario 4 in the open-field environment,
i.e., the NLoS scenario, using a window size of 10 λ.
In what follows, the spatial correlation of the SF for Scenario 1 and Scenario
4 of the open-field environment is investigated. Scenario 4 is further divided into
a LoS part (parked vehicles are not obstructing the LoS) and an OLoS part (six
parked vehicles are obstructing the LoS). The LoS part covers the Tx-Rx distance
larger than 60 m, while the OLoS part covers the Tx-Rx distance less than 60 m.
The following analysis is published in [16].
The spatial correlation can be obtained by the autocorrelation of the SF as:

r(∆d) = E [X(d)X(d + ∆d)] , (4.10)

48
 4.2 Spatial Correlation of Shadow Fading in Open-field

100

80

60 Meas pathloss
Loss [dB]

Log-distance pathloss
40 Shadow fading
Small-scale fading
20

-20
101 102
Tx-Rx distance [m]
Figure 4.3: An example of the extraction of the SF from the measured path loss for
Scenario 4 in the open-field using a window size of 10λ.

where d is the distance between the Tx and the Rx, and ∆d is the distance
between two observed positions. E [·] denotes the expected value of [·]. In principle,
the estimation of the autocorrelation function of any random process requires having
a large set of samples at every observation time t to calculate the statistical average
or the so called ensemble average. By assuming ergodicity of a random process, i.e.,
the sample average over time t for one realization (measurement run) of the random
process converges to the ensemble average as the length of the realizations tends to
infinity, the autocorrelation function can be estimated by using samples collected
from one measurement run.

A random process is called WSS when the mean and the autocorrelation do
not vary with time. The SF can be assumed WSS since the distance-dependent part
is subtracted from the measured path loss when extracting the SF [60]. However,
to prove the WSS assumption, the SF is put under test, namely, the modified
reverse arrangement test (MRA) [61]. The MRA test is performed by dividing the
SF sample record into Ns equal segments and calculating the square mean value

49
4. Channel Characteristics and General Modeling Aspects

in each segment (x21 , x22 , x23 , . . . x2N s ). In a stationary random process the square
mean values of adjacent segments are independent [61] and any time trend will
result in non-stationarity. The test checks if the examined random process has a
time trend by calculating how many times, starting with x21 , that each subsequent
square mean value (x22 , x23 , . . . x2N s ) is less than x21 . This step is repeated with each
square mean value. Each inequality is called a ’reverse arrangement’. Figure 4.4
shows an example of the mean square values of the SF for Ns =10 segments and
Ns =20 segments which correspond to distance interval of 5 m and 10 m for both
Scenario 1 and Scenario 4, respectively. It can be seen that the changes in the mean
value do not follow a specific trend. The total number of the reverse arrangements
A is then used to calculate the total score as
[︂ ]︂
Ns (Ns −1)
A− 4
z = √︂ . (4.11)
2Ns3 +3Ns2 −5Ns
72

100
5m interval-Scenario 4
10m interval-Scenario 4
80 5m interval-Scenario 1
Mean square of SF [dB]

10m interval-Scenario 1

60

40

20

0
0 5 10 15 20
Interval index

Figure 4.4: Mean square value of the SF calculated within intervals of 5 m and 10 m
for Scenario 1 and Scenario 4 in open-field.

The stationarity hypothesis is then verified with 5 % significance if the absolute

value of the total score −1.96 ≤ z ≤ 1.96. Figure 4.5 shows the z score of the

50
 4.2 Spatial Correlation of Shadow Fading in Open-field

MRA test for 60 different number of intervals Ns . It can be seen that both scenarios
as well as the two parts of the Scenario 4 pass the stationarity test.

2 2
Scenario 1 Scenario 4 - LoS part
1.5 Scenario 4 1.5 Scenario 4 - OLoS part

1 1

0.5 0.5
Test score

Test score
0 0

-0.5 -0.5

-1 -1

-1.5 -1.5

-2 -2
0 5 10 15 20 25 0 2 4 6 8 10 12
Interval length [m] Interval length [m]
(a) Scenario 1 and Scenario 4 (b) LoS and OLoS parts of Scenario 4

Figure 4.5: Score of the MRA stationarity test.

The autocorrelation functions is estimated from the measurement data using

Equation (4.10). Figure 4.6 shows the estimated autocorrelation functions with
respect to ∆d.
The correlation distance dc is defined as the value of ∆d at which the value of
the normalized autocorrelation function drops to e−1 . The correlation distance is
scenario-dependent and indicates how fast the SF is changing over distance. The SF
can be assumed to stay constant within the correlation distance dc . Applying this
assumption, the SF can be considered as a block fading which leads to a simplification
of modeling the SF in simulators.
Figure 4.6a illustrates the spatial autocorrelation function of the SF for both
scenarios. It can be seen that the correlation distance in Scenario 1 is 7.2 m. Blocking
the LoS by parked vehicles in Scenario 4 leads to a larger correlation distance of
10.4 m. This implies that compared to the LoS, the SF in the NLoS experiences less
variation, i.e., similar diffraction loss, when the LoS is obstructed by the same object.
The oscillation pattern in both scenarios can be explained by the constructive and
destructive interference, which is due to the superposition of the direct and the
ground-reflected rays, (see Figure 4.1a and Figure 4.1d).

51
4. Channel Characteristics and General Modeling Aspects

1 1
Scenario 1 Scenario 4 - LoS part
0.8 Scenario 4 Scenario 4 - OLoS part

0.6
r ( (d))/r (0)

0.5

r x( (d))/rx(0)
x

0.4

0.2
x

0
0

-0.2

-0.4 -0.5
0 20 40 60 80 100 120 0 10 20 30 40 50 60
d [m] d [m]
(a) Scenario 1 and Scenario 4 (b) LoS part and the OLoS part of Scenario 4

Figure 4.6: Normalized empirical spatial autocorrelation function of the SF for

Scenario 1, Scenario 4 and the two parts of Scenario 4 in open-field.

The spatial autocorrelation function of the SF for both LoS and OLoS parts of
Scenario 4 are illustrated in Figure 4.6b. Both the LoS and the OLoS parts have
nearly similar correlation distances of 10.7 m and 10.9 m, respectively. However, the
SF in the OLoS part experiences less correlation than the LoS part when ∆d < 10 m.
This indicates that the SF varies faster compared to the LoS part within ∆d < 10 m,
as can be seen in Figure 4.3. Moreover, the oscillation pattern observed in the OLoS
part in Figure 4.6b is probably due to the blockage of the LoS by parked vehicles.
The empirical autocorrelation function is fitted to three theoretical autocorrela-
tion function models, the first model is the classical model proposed by Gudmundson
[62], which is based on a single negative exponential function,
(︄ )︄
|∆d|
r(∆d) = exp − , (4.12)
dc1

where dc1 is the correlation distance. The second model is the double exponential
model, which is also a well-known and widely used model [63]. It models the spatial
correlation as a sum of two negative independent exponential functions,

∆d ∆d
(︄ )︄ (︄ )︄
r(∆d) = α exp − + (1 − α) exp − , (4.13)
dc2 dc3

52
 4.2 Spatial Correlation of Shadow Fading in Open-field

where dc2 > 0, dc3 > 0 and the weight factor 0 ≤ α ≤ 1 are tunable parameters.
The third model is the exponential decaying sinusoid model [64] and is given by

∆d ∆d ∆d
(︄ )︄ [︄ (︄ )︄ (︄ )︄]︄
dc5
r(∆d) = exp − cos − + sin − , (4.14)
dc4 dc5 dc4 dc5

where dc4 > 0 and dc5 > 0 are tunable parameters. The tunable parameters in
Equations (4.13) and (4.14) are estimated in a MMSE sense.

1 1
Empirical Empirical
0.8 Single exponential Single exponential
Double exponential Double exponential
0.6 Sinusoid Sinusoid
r x( (d))/rx(0)

r x( (d))/rx(0)

0.5
0.4

0.2
0
0

-0.2

-0.4 -0.5
0 20 40 60 80 100 120 0 20 40 60 80 100 120
d [m] d [m]
(a) Scenario 1 (b) Scenario 4

Figure 4.7: Normalized empirical spatial autocorrelation function of the SF and the
corresponding fitting models for Scenario 1 and Scenario 4 in open-field.

From Figure 4.7 and Figure 4.8 and Table 4.3, it can be found that the single
and the double exponential model can loosely follow the trend of the empirical
autocorrelation function. The double exponential model performs slightly better
than the exponential model with two parameters to be estimated rather than one.
The exponential decaying sinusoid model provides better match to the empirical
autocorrelation function in all scenarios except for the OLoS part in Figure 4.8b
where all three models provide a fit with a nearly similar standard deviation σ.

53
4. Channel Characteristics and General Modeling Aspects

1 1
Empirical Empirical
Single exponential Single exponential
Double exponential Double exponential
Sinusoid Sinusoid
0.5 0.5
r x( (d))/rx(0)

r x( (d))/rx(0)
0 0

-0.5 -0.5
0 10 20 30 40 50 60 0 10 20 30 40 50
d [m] d [m]
(a) LoS part (No vehicles are obstructing the (b) OLoS part (six vehicles are obstructing
LoS). the LoS).

Figure 4.8: Normalized empirical spatial autocorrelation function of the SF and the
corresponding fitting models for the LoS and OLoS parts of Scenario 4 in open-field.

Scenario 4
Model Par Scenario 1 Scenario 4
OLoS part LoS part
Single dc1 [m] 7.2 10.4 10.7 10.9
exponential σ 0.12 0.13 0.083 0.13
dc2 [m] 5 7 1 8
Double dc3 [m] 5 7 7 8
exponential α 0.1 0.1 0.13 0.1
σ 0.11 0.11 0.080 0.11
dc4 [m] 16 30 3 23
Sinusoid dc5 [m] 6 8 46 9
σ 0.04 0.03 0.084 0.03

Table 4.3: Spatial correlation models’ parameters.

54
 4.3 Diffraction Loss Model

4.3 Diffraction Loss Model

When the LoS path between the Tx and the Rx is obstructed in the Fresnel zone
by an obstacle whose dimensions are larger than the wavelength of the radio wave,
the measured propagation path loss is increased. The additional increase in the
attenuation is due to the blockage of the LoS by the obstacles and the signal is
received by diffraction of the electromagnetic waves. According to Huygens principle,
the electric field is the sum of the Huygens sources located in the plane above the
obstruction [30]. The calculation of the diffraction loss can be done by treating the
obstacles as absorbing knife-edges [30]. Applying this simplification, the diffraction
loss (in dB) becomes a function of only the Fresnel-Kirchoff parameter v as:
⎛ √︂ ⎞
[1 − C(v) − S(v)]2 + [C(v) − S(v), ]2
Li = −20 log ⎝ (4.15)
2
⎠

where C(v) and S(v) are the Fresnel cosine and sine integrals.
The complex Fresnel integral is given by:
(︄ )︄
∫︂ v
πt2
F (v) = exp j dt = C(v) + jS(v), (4.16)
0 2

The Fresnel cosine and sine integrals are given by:

(︄ )︄
∫︂ v
πt2
C(v) = cos dt (4.17)
0 2
(︄ )︄
∫︂ v
πt2
S(v) = sin dt (4.18)
0 2
The sine and cosine integrals can be calculated by using their Taylor expansion as
in [65],
∞
C(v) = cm v 4m+1 , c0 = 1,
∑︂

m=0
(4.19)
−π 2 (4m + 1)cm
cm+1 = .
4(2m + 1)(2m + 2)(4m + 5)

55
4. Channel Characteristics and General Modeling Aspects

∞
π
S(v) = sm v 4m+3 , s0 =
∑︂
,
6
m=0
(4.20)
−π 2 (4m + 3)sm
sm+1 = .
4(2m + 2)(2m + 3)(4m + 7)
The Fresnel-Kirchoff parameter v depends on the distance d1 from the Tx to
the diffracting edge, the distance d2 from the diffracting edge to the Rx and the
effective height h of the diffracting edge:
√︄
2(d1 + d2 )
v=h . (4.21)
λd1 d2
To calculate the diffraction loss over multiple knife edges, the Epstein-Peterson
method is used [66] which is illustrated in Figure 4.9. This method is found to give
the best results for the given geometry. In this method, the total diffraction loss is
the sum of the k losses on all edges:

k
Ltot = (4.22)
∑︂
Li .
i=1

Figure 4.9: The Epstein-Peterson method for four edges. For the edge B, the
geometrical parameters (d1 , d2 , h) in Equation (4.21) are (d1 , d2 , hB ). Similarly, for
edge C, the geometrical parameters (d1 , d2 , h) in Equation (4.21) are (d2 , d3 , hC ).
This figure can be seen as a 2D projection of the 3D model in Figure 4.11.

56
 4.3 Diffraction Loss Model

When calculating the loss due to diffraction from the first edge, the second edge
is considered as a receiver and then to calculate the loss on the second edge, the first
edge is considered as a transmitter and the third edge as a receiver and so forth.

Figure 4.10: Snapshot from the 3D modeling too where the Tx (black vehicle) is
shown at different positions, and the Rx (Pedestrian) is shadowed by parked vehicles.

Hence, modeling the diffraction loss requires computing the geometrical

parameters d1 , d2 and h for the diffraction process at each edge. This can be
done by accurately modeling the size of the objects and their distribution in the
propagation environment. To this end, a 3D tool, illustrated in Figure 4.10, is
developed and published in [16]. The modeling procedure comprises the following
steps:

1. Modeling static objects in the environment is done by importing the objects

dimensions and locations from the laser scanner measurements and the positions
of Tx and Rx from the GNSS measurements. The vehicles are represented by
a cuboid and the vehicle front-end is represented by a triangular prism.

2. A direct ray from the Tx to the Rx is created, then intersections between

this ray and the objects in the environment are detected as illustrated in
Figure 4.11.

3. If the ray intersects with an object, two intersection points will result, entering
point and exiting point. The diffraction points are found by projecting the

57
4. Channel Characteristics and General Modeling Aspects

Figure 4.11: 3D illustration of the modeling procedure.

intersection points on the roof of the object. For simplicity, only roof diffraction
is considered.

4. The exiting point is then treated as a secondary transmitter and the previous
step is repeated to check further intersections with other objects and find the
other diffraction points.

5. The diffraction loss is then calculated at each diffraction point and summed
up to get the total diffraction loss.

6. The diffraction loss is combined with the two-ray path loss model to get the
total propagation loss.

In Figure 4.12, the measured path loss within the obstruction region in Scenario
4 in the open-field is compared with the modeled path loss using the combined
two-ray path loss model and the multiple knife-edge diffraction model. Despite its
simplicity, the proposed model is able to provide a good match for the path loss with
a standard deviation σ = 4 dB when considering a line of parked vehicles between a
vehicle and a pedestrian. The discontinuity of the model’s curve occurs when the
number of parked vehicles blocking the LoS is changed due to the movement which
will result in a sudden increase or decrease in the number of diffraction edge.

58
 4.3 Diffraction Loss Model

Figure 4.12: The measured path loss during obstruction of the LoS, and the modeled
path loss as a summation of the multiple knife-edge diffraction model and the two-ray
path loss model for Scenario 4 in open-field. The measured path loss is plotted in
different colors where each color corresponds to the obstruction of the LoS by a specific
vehicle. For example, the path loss in red contains the diffraction loss caused by the
vehicle near the collision point.

59
4. Channel Characteristics and General Modeling Aspects

4.4 Non-stationarity Analysis

WINNER-type GSCM channel models [15] require estimation of the LSPs and
their correlation properties, such as the SF, KF, DS, and AS. For simplification,
the time-variant channels are usually treated as WSSUS. Basically, a channel is
called WSSUS when the first and second moments are independent from the center
frequency and absolute time. However, the WSSUS assumption is often not valid
especially in vehicular communications due to the high mobility of vehicles, VRUs,
and the surrounding scatterers. Consequently, it is particularly important due to the
non-stationarity to characterize the channel assuming local quasi-stationarity (LQS)
within small regions or segments in time and frequency. The channel is divided
into regions where the WSSUS assumption holds and then LSPs are estimated in
each individual region. Estimation of the size of the LQS regions is important not
only for the estimation of the LSPs but also for understanding the evolution of
the propagation channel. The stationarity of the V2VRU channel is assisted by
estimating the GLSF and its collinearity based on the channel measurement data as
proposed in [67].
As detailed in [68], the local scattering function (LSF) is defined for non-WSSUS
channel in continuous case as

{︂ }︂
CH (t, f ; ν, τ ) ≜ F∆t→ν F−1
∆f →τ {RH (t, f ; ∆t, ∆f )} , (4.23)

where
RH (t, f ; ∆t, ∆f ) ≜ E{LH (t, f + ∆f )L∗H (t − ∆t, f )} (4.24)

is the 4D correlation function, LH (t, f ) is the time-varying channel transfer function,

and E{.} denotes the mathematical expectation. However, the LSF is not always
positive and it depends on the whole correlation function RH (t, f ; ∆t, ∆f ). Therefore
[68] defines a smooth version of the LSF which is the GLSF as

(Φ)
CH (t, f ; ν, τ ) ≜ (CH∗4 Φ)(t, f ; ν, τ ), (4.25)

60
 4.4 Non-stationarity Analysis

where
K ∫︂ ∞ ∫︂ ∞
Φ(t,f ν, τ ) = L∗Gk (−t, −f + ∆f )
∑︂
γk
k=1 −∞ −∞ (4.26)
× LGk (−t − ∆t, −f )e−j2π(ν∆t−τ ∆f ) d∆t d∆f,

where LGk (t, f ) are windowing functions in time-frequency, γk ≥ 0 are normalizing

constants that sum up to one, and K is the number of used windows .
Since the recorded channel transfer functions are available at discrete time and
frequency instants, the discrete representation of the GLSF is given by [69]

K−1 {︂ }︂
(Φ)
CH [m, q; p, n] = γk E |H Gk [m, q; p, n]|2 (4.27)
∑︂
.
k=0

The tapered frequency response is

√︂ ⌈Nt /2⌉−1 ⌈Nf /2⌉−1

Gk
[m, q; p, n] = L∗Gk [m′ , q ′ ]
∑︂ ∑︂
H Ts Fs
m′ =−⌊Nt /2⌋ q ′ =−⌊Nf /2⌋ (4.28)
(︂ )︂
pm′ ′
−j2π − ng
× LH [m + m′ , q + q ′ ]e Nt Nf
.

LH [m, q], LGk [m, q], m, p, q, and n are the discrete representations of the
continuous case, Nt and Nf are the time and frequency windows length, Ts and Fs
are the difference in successive samples in time and frequency.
To evaluate the stationarity time, the collinearity between different time instants
of the GLSF is estimated as
{︃ }︃
(Φ)H (Φ)
tr CH [m, q]CH [m′ , q]
α[m, m′ ] ≜ ⃦⃦ (Φ) ⃦ ⃦
⃦ ⃦ (Φ)
⃦ , (4.29)
⃦CH [m, q]⃦ ⃦CH [m′ , q]⃦
⃦
F F

where tr indicates a trace operation and ∥.∥F refers to frobenius norm. The resulting
collinearity matrix is squared and symmetric, and the main diagonal contains values
of 1. The collinearity α[m, m′ ] ∈ [0, 1], where 1 means a maximum correlation and 0
corresponds to no correlation. The collinearity between two GLSFs will have high
value when the overlap between them is significant, and vice versa.

61
4. Channel Characteristics and General Modeling Aspects

After the collinearity is estimated, sets Mj [m] are defined using a collinearity
threshold thcol as
Mcol [m] ≜ {m′ |α[m, m′ ] > thcol } . (4.30)

Finally, the time dependent stationarity time is obtained

Tstat,j [m] ≜ |Cj [m]|Ts (4.31)

where Cj [m] is a connected subset of Mj [m] that has a maximum cardinality and
Ts is the samples time spacing.
To estimate the GLSF, a sliding window mechanism, with one snapshot shift, is
used. A discrete prolate spheroidal sequence (DPSS) [70] is employed as a window
function. Four DPS sequences in time domain and two in frequency domain (I = 4
and J = 2) are used, thus leading to a total of K = 8 2D windows. The window
length in time domain is set to Nt = 30 snapshots, which corresponds to 0.03 s
or approximately 0.2 m. Since the analysis is conducted for a fixed frequency, the
frequency dependency on the GLSF is omitted and only the time dependency is
considered. Therefore, the window length in frequency domain is set to Nf = 384
frequency bins, which corresponds to the total bandwidth. The collinearity is then
calculated between all GLSF delay x Doppler elements. The stationarity distance is
estimated assuming a threshold thcol = 0.9 and taking into account the speed of the
vehicle.
Figure 4.13 shows an example of the estimated GLSF for Scenario 3 in the urban
environment at four different time instances. In Figure 4.13a, t = 1 s at Tx-Rx
distance of 93 m, the vehicle is at the beginning of the track and moving towards
the collision point, the pedestrian is standing still and shadowed by nearby parked
vehicles. The LoS is diffracted at the roof of the vehicle and has a delay of 0.31 µs.
Three MPCs with positive Doppler frequency are located nearby the LoS with delays
between 0.32 µs and 0.34 µs originated from a building and parked vehicles on the
road side opposite to the pedestrian. A negative Doppler MPC at a delay of 0.36 µs
can also be seen. This MPC is due to a reflection by a building behind the vehicle.
In Figure 4.13b, t = 5 s at distance of 55 m, the delay of the LoS component is
decreased to approximately 0.18 µs as the vehicle moves towards the collision point.

62
 4.4 Non-stationarity Analysis

The most significant MPCs have a delay between 0.19 µs and 0.26 µs with positive
Doppler frequencies. These MPCs are originated from parked vehicles and buildings
on both sides of the road. At distances between 55 m and 70 m, the LoS becomes less
obstructed and it experiences less power fluctuations which will later has a major
impact on the stationarity distance.

(a) t = 1 s (b) t = 5 s

(c) t = 7 s (d) t = 11 s

Figure 4.13: Examples of the estimated GLSF at different time instants for Scenario
3 in the urban environment.

In Figure 4.13c, t = 7 s at Tx-Rx distance of 42 m, the pedestrian is attempting

to cross the road and moving towards the collision point. Due to this movement,
the LoS starts to experience rapid variations in power. One main difference to

63
4. Channel Characteristics and General Modeling Aspects

the previous time instant is the new MPC with zero Doppler frequency at delay
of 0.26 µs which originated by reflections from parked vehicles located nearby and
parallel to the moving vehicle. In Figure 4.13d, t = 11 s at Tx-Rx distance of 7 m,
the vehicle is approximately 1 m far from the collision point where a strong LoS at
0.023 µs can be observed followed by three MPCs induced by nearby vehicles and
building.
Figure 4.14 presents examples of the collinearity of the GLSF for the three
scenarios in the urban environment. In all scenarios the GLSF at different Tx-Rx
distances is correlated only for short travel distance/time which is implied by the
observed decrease of collinearity away from the main diagonal. This decrease can
be explained by the decrease of the correlation between two GLSFs as the time
differences between them increases. In other words, the two GLSFs experience high
correlation only for a short time period.
The collinearity of Scenario 1 in Figure 4.14a decreases much faster than in
Scenario 2 and Scenario 3. This decrease can be explained by the transit from LoS
to NLoS at around 30 m whereas in the other two scenarios the OLoS situation is
maintained throughout most of the experiment with an exception at the collision
point. The time variant structure of the collinearity in all scenarios is confirming
the non-stationary nature of the propagation channel.
The estimated stationarity time is mapped to stationarity distance using the
position information of the Tx and Rx. Figure 4.15 shows the corresponding
stationarity distance to the collinearity of the GLSF depicted in Figure 4.14.
In all scenarios, at the shortest Tx-Rx distance, i.e., at the collision point, the
stationarity distance suddenly drops. This results from the sudden change in the
Doppler domain due to the LoS component, which shifts from positive to negative
Doppler at the moment of passing the collision point.
Figure 4.15a shows the estimated stationarity distance for Scenario 1. An
increase in the stationarity distance is observed at distances less than 20 m due
to the appearance of the LoS and several strong reflections. In NLoS situation,
when the cyclist is blocked by buildings, the stationarity distance varies between
0.5 m and 2.8 m. The peaks at 35 m and 70 m coincide with the appearance of a

64
 4.4 Non-stationarity Analysis

90 1 1
90
80 80
0.8 0.8
70 70
Tx-Rx distance [m]

Tx-Rx distance [m]

60 0.6
60 0.6
50 50

40 0.4 40 0.4
30
30
0.2 20 0.2
20
10
10
10 20 30 40 50 60 70 80 90 10 20 30 40 50 60 70 80 90
Tx-Rx distance [m] Tx-Rx distance [m]
(a) Scenario 1 (b) Scenario 2

1
90
80 0.8
Tx-Rx distance [m]

70
60 0.6

50
40 0.4

30
20 0.2

10

10 20 30 40 50 60 70 80 90
Tx-Rx distance [m]
(c) Scenario 3

Figure 4.14: Examples of collinearity of the GLSF in the urban environment.

65
4. Channel Characteristics and General Modeling Aspects

strong reflection from nearby building and a diffracted LoS on the building corner,
respectively. The average stationarity distance dstat in Scenario 1 is 2.13 m.
In Scenario 2 (Figure 4.15b), where the cyclist and the vehicle move on parallel
tracks with an OLoS situation during the whole scenario, the average stationarity
distance dstat increases to 5.25 m. The maximum stationarity distance during the
whole scenario is around 7 m except near the collision point where the stationarity
distance rapidly varies between 3 m and 21 m due to the strong variation in LoS
power. A pattern of consecutive highs and lows is observed which can be explained
by the appearance and disappearance of the LoS due to the blockage by the parked
vehicles. Generally, stronger LoS component leads to larger stationarity distance.
A similar pattern is observed again in Scenario 3 in Figure 4.15c. The stationarity
distances in Scenarios 2 and 3 are within similar range. However, in Scenario 3 the
stationarity distance at the collision point is 5 m, which is shorter than in Scenario 1
and Scenario 2. By analyzing the scatterers and their locations in Section 4.6, less
reflections are found at the collision point in Scenario 3 compared to Scenario 1 and
Scenario 2 which could be the reason behind the shorter stationarity distance. The
average stationarity distance dstat in this scenario is 3.58 m. Table 4.4 summarizes
the average stationarity distances for all scenarios.

Scenario 1 Scenario 2 Scenario 3

dstat [m] 2.13 5.25 3.58

Table 4.4: Average stationarity distances in the urban environment.

The aforementioned comparison is depicted in Figure 4.16, which shows the

cumulative distribution functions (CDFs) of the stationarity distances for all
scenarios. Note that the the CDFs are plotted for each scenario from all measurement
runs.

Stationarity Distance Relation with LoS Component, and KF in Scenario

3

To get more insight into the stationarity distance, the influence of the LoS component
power on the stationarity distance is investigated. All MPCs are filtered out from

66
 4.4 Non-stationarity Analysis

7 25

6
20
Stationarity distance [m]

Stationarity distance [m]

5

4 15

3 10
2
5
1

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Tx-Rx distance [m] Tx-Rx distance [m]
(a) Scenario 1 (b) Scenario 2

5
Stationarity distance [m]

0
0 20 40 60 80 100
Tx-Rx distance [m]
(c) Scenario 3

Figure 4.15: Stationarity distances in urban environment.

67
4. Channel Characteristics and General Modeling Aspects

1
Scenario 1
Scenario 2
0.8 Scenario 3

0.6
CDF

0.4

0.2

0
0 5 10 15 20 25 30
Stationarity distance [m]
Figure 4.16: CDFs of the stationarity distances for all scenarios in the urban
environment.

the corresponding CIR of Scenario 3 and only the LoS component remains. The
GLSF is then estimated and based on its collinearity, the stationarity distance is
estimated.
The received power carried by the LoS is plotted alongside the power carried
by all MPCs, including the LoS, in Figure 4.17a. It can be seen that the total
received power increases from approximately −60 dBm when the distance between
the vehicle and the pedestrian is 100 m to −20 dBm at the collision point due to a
strong LoS. The total received power and the power carried by the LoS component
experience rapid fluctuations when the LoS between the vehicle and the pedestrian
is partially or completely obstructed by parked vehicles. At a distance between 55 m
and 70 m the LoS becomes less obstructed due to large gap between the parked
vehicles, therefore, a slight increase in the received power and less fluctuations are
observed. In Figure 4.17b the estimated K-factor is shown as it provides valuable
information about the contribution of the LoS to the total received power and hence
its impact on the stationarity distance. It is found that the K-factor mostly has
positive values and it reaches 31 dB near the collision point. Due to diffraction

68
 4.4 Non-stationarity Analysis

loss, the LoS at some distances loses up to 15 dB from its power and therefore the
K-factor drops to approximately −2 dB. However, this drop lasts only for short
period.
The collinearity of the GLSF, when only the LoS component exists, is plotted
in Figure 4.17c. As expected, the collinearity of the GLSF has low values at
distances where the LoS undergoes rapid fluctuations, which will result in short
stationarity distance. Figure 4.17d shows a comparison of the estimated stationarity
distance in Scenario 3 when all MPCs are accounted for and when only the LoS
component is considered. As seen from the K-factor, the LoS component is the most
dominant component, therefore, the power of the LoS has the main effect on the
stationarity distance. Figure 4.17b shows that the higher the K-factor, indicating
higher dominance of the LoS on the total received power, the larger the stationarity
distance. However, not only the power of the LoS but also its variation affects the
stationarity distance. As can be noticed from Figure 4.17a, the larger and more rapid
the power variations of the LoS component the shorter the stationarity distance.

69
4. Channel Characteristics and General Modeling Aspects

-20 35
LoS component
All MPCs 30
-30
Received power [dBm]

25
-40

K-factor [dB]
20

-50 15

10
-60
5
-70
0

-80 -5
0 20 40 60 80 100 0 20 40 60 80 100
Tx-Rx distance [m] Tx-Rx distance [m]
(a) Received power (b) K-factor

8
1
LoS component
90 7 All MPCs
Stationarity distance [m]

80 0.8
6
Tx-Rx distance [m]

70
5
60 0.6
4
50
40 0.4 3
30 2
20 0.2
1
10
0
10 20 30 40 50 60 70 80 90 0 20 40 60 80 100
Tx-Rx distance [m] Tx-Rx distance [m]
(c) Collinearity of GLSF for only LoS compo- (d) Stationarity distance
nent

Figure 4.17: Stationarity distance relation with the power of the LoS component
and the KF in Scenario 3.

70
 4.5 MPCs Parameters Estimation and Tracking

4.5 MPCs Parameters Estimation and Tracking

The raw measurement data obtained from channel sounding measurements cannot
be directly used for a detailed characterization of the propagation channel when
a GSCM channel modeling approach is targeted. Therefore, a super resolution
estimation algorithm can be employed to estimate the parameters of the MPCs.
Few algorithms can be found in the literature like the snapshot-based RIMAX
[71], and the space-alternating generalized expectation-maximization (SAGE) [72].
Alternatively, an algorithm that takes the evolution of the multipath overtime into
account such as KEST [73] will be used in this thesis.
Examples of the measured CIRs in the open-field and urban environments are
shown in Figure 4.18. At the beginning of the experiment, the Tx-Rx separation
distance was approximately 100 m. The LoS between the Tx and Rx was maintained
in the open-field Scenario 1 with some traces parallel to the LoS. In the urban
Scenario 1, as can be seen in Figure 4.18b, the LoS starts to appear, accompanied
by a group of strong multipaths, at approximately 40 m distance or 12.5 s. The LoS
path remains between 12.5 and 19 s but is partially obstructed by parked vehicles.
During NLoS, a weak path appears as a tail of the LoS path and with a slightly
more delay than the geometric LoS (GLoS) accompanied with a group of weak paths.
Contrary to the open-filed, in the urban measurements, the CIR is highly cluttered.
This clutter mainly consists of DMCs, but it could also contain artifacts, noise,
and unreliable specular reflections, which have short lifetime (in the order of a few
wavelengths) or weak power.
The behavior of the multipath channel can be described by the time variant CIR
h(tn , τ ), which can be expressed mathematically as a sum of N (tn ) Dirac impulses
[55]:
N (tn )−1
h(tn , τ ) = αm (tn ) · δm (τ − τm (tn )), (4.32)
∑︂

m=0

where αm (tn ), τm (tn ) denote the complex amplitude and the delay of the mth MPC
in snapshot n, and δ(·) is a Dirac distribution. A dynamic multipath estimator
named KEST, introduced in [73], is employed for estimating the parameters of SMCs.
The estimated parameters are the absolute value of the amplitude, the delay, and the

71
4. Channel Characteristics and General Modeling Aspects

0.8 -10 0.8 -10

0.7 -15 0.7 -15

0.6 -20 0.6 -20

0.5 -25
Delay [ s]

0.5 -25

Delay [ s]
Power [dB]

Power [dB]
0.4 -30 0.4 -30

0.3 -35 0.3 -35

0.2 -40 0.2 -40

0.1 -45 0.1 -45

0 -50 0 -50
0 2 4 6 8 10 0 5 10 15
Time [s] Time [s]
(a) Open-field: Scenario 1 (b) Urban: Scenario 1

Figure 4.18: Examples of the time-variant CIRs based on the measured data.

phase of each MPC at time instant tn . KEST uses the output of SAGE algorithm
[72] as measurements within a Kalman filter. SAGE is used as a snapshot-based
estimator that jointly estimates the complex amplitude αm (tn ) and the delay τm (tn )
for each MPC m. Additionally, KEST consists of several Kalman filters in parallel
using different model orders for estimating the number of MPCs. Figure 4.19 shows
the CIRs for open-field and urban scenarios based on the estimated parameters by
the KEST algorithm. The signal period Tp of the periodically transmitted multi-
tone signal in the open-field environment was selected to be equal to 0.8 µs, which
corresponds to a propagation distance of 240 m. Therefore, reflections with larger
propagation distance were superimposed with the CIR of the next measurement
snapshot. Nevertheless, the delays of these MPCs are corrected by adding 0.8 µs in
the upcoming analyses. In the open-field environment, the LoS between the Tx and
the Rx was never obstructed and, consequently, the first measured path coincides
with the geometric LoS. However, in the urban environment, from 0 to 12.5 s the
LoS was continuously obstructed by buildings. Since the first path reaches the Rx
after diffracting at the edge of the corner building, there is a misalignment between
the first arriving path and the GLoS. At approximately 12.5 s or 40 m distance, the
LoS path appears. A good alignment between the estimated LoS path and the GLoS
can be noticed.

72
 4.5 MPCs Parameters Estimation and Tracking

It can be seen that the estimation results in both scenarios confirm the initial
observations made on the measured CIRs in Figure 4.18. Based on the time-delay
evolution structure of the MPCs, it can be deduced, based on the long lifetimes,
that all estimated MPCs in the open-field scenario are due to specular reflections.
However, in the urban scenario, not only SMCs but also DMCs can be seen. Therefore,
extracting the SMCs from the estimated time-variant CIR is essential in order to
model the SMCs and DMCs separately.

(a) Open-field: Scenario 1 (b) Urban: Scenario 1

Figure 4.19: Time-variant CIRs based on the estimated parameters. The geometric
LoS is displayed as a red dashed line.

4.5.1 Specular Multipath Extraction

MPCs can be divided into specular reflections (SMC) and diffuse scattering (DMC).
Paths with relatively high power and long lifetime are considered as SMCs. SMCs,
on one hand, are coherent, which means that they have a relatively constant phase
difference with the LoS path over time. While, on the other hand, DMCs will have
random amplitudes and phases and therefore they are called incoherent components.
Unlike SMCs, DMCs have no clear structure in the time-delay domain and are
often modeled as colored noise [74, 71]. Mirror-like smooth surfaces lead to only
SMCs, which are also referred to as smooth reflections. Mirror-like surfaces are
rarely found in outdoor environments where most surfaces possess some roughness

73
 4. Channel Characteristics and General Modeling Aspects

(roughness in relation with the wavelength). The roughness leads to a decrease

in the amplitude of the reflected path. This reduction is due to the scattering of
energy in different directions. This scattered energy is carried by DMC. DMC is
scattered from all or part of the rough surface and this part is called a glistening zone
[29]. A very rough surface may result in canceling the contribution of the diffuse
(incoherent) components. However, with a less rough surface, the DMCs contribute
to the received signal and together with the specular (coherent) components produce
an interference pattern on the path’s amplitude. Similar to the rough surfaces,
vegetation could also produces both SMCs and DMCs [75, 76].
As seen in Figure 4.19b, in the urban environment, the output of the MPCs
parameters estimation not only contains SMCs but also contains DMCs. For channel
modeling, the dominant SMCs with long lifetime need to be extracted. Therefore, a
method to separate SMCs and DMCs based on their number of neighboring MPCs
that have delay difference less or equal a specific threshold is proposed and published
in [17]. The steps of the proposed method are summarized as follows:

0. For each MPC m in measurement snapshot n, a search radius δ is defined. The

search region has a length of 2δ, and includes all snapshots with indices n + p,
where p = [−δ, . . . , −1, 1, 2, . . . , δ].

1. Check if the MPCs within the search region are close in delay as
⎧
⎨1
⎪
if |τm,n − τm̃,n+p | ≤ ζ∆τ
Tm̃,n+p = (4.33)
⎩0
⎪
otherwise,

where τm,n is the delay of the MPC m in snapshot n at time instant tn , τm̃,n+p is
the delay of the MPC m̃ in snapshot n + p, ζ∆τ is the delay difference threshold,
"1" means the MPC under examination is close in delay, and "0" means it is not.
The set T is defined as
⃓
T = {Tm̃,n+p ⃓ p = −δ, . . . , −1, 1, 2, . . . , δ, m̃ = 1, 2, . . . , Nn+p }, (4.34)
⃓

where Nn+p is the number of MPCs in snapshot n + p.

74
 4.5 MPCs Parameters Estimation and Tracking

2. The MPC is classified as specular or diffuse multipath as

⎧
⎨Specular if
∑︂
⎪
⎪ Tm̃,n+p ≥ ζL ,
MPC =
ˆ︁ Tm̃,n+p ∈T (4.35)
⎩Diffuse otherwise,
⎪
⎪

where ζL is a threshold on the number of neighbors. The MPC is considered

as specular if the number of detected MPCs in the region tn−δ , . . . , tn+δ and
τm,n − ζ∆τ , . . . , τm,n + ζ∆τ is larger than or equal the threshold ζL .

Figure 4.20 shows an example of using this method on a small part of the CIR
shown in Figure 4.19b. Two MPCs, are marked by red circles, at t =17.38 s, and
the search region for each of them is highlighted by a red rectangle. Based on the
number of neighboring MPCs within the search region, the MPC at around 0.13 µs is
classified as an SMC while the MPC at 0.11 µs is classified as a DMC and therefore
it is filtered out.
Parameter Name Value
δ Search region radius 25 snapshots
ζ∆τ Delay difference threshold 1.5 ns
ζL Nr. of neighbors threshold 10 MPCs

Table 4.5: Specular reflections extraction parameters.

The values of the thresholds ζL and ζ∆τ depend on the richness of the DMCs.
These values affect the performance of the extraction and therefore should be
carefully chosen. By decreasing ζ∆τ more SMCs with high delay variations will
be mistakenly classified as DMCs. Contrarily, by increasing ζ∆τ more DMCs will
be mistakenly classified as SMCs. In order to compare the effect of variations in
the ζ∆τ parameter, SMCs extraction is performed with three different ζ∆τ values.
Figure 4.21 only shows the extracted SMCs for the part of the CIR within the red
rectangle depicted in Figure 4.19b. With larger ζ∆τ more MPCs are classified as
SMCs.
Similarly, decreasing ζL will result in more DMCs that are mistakenly classified
as SMCs. On the other hand, when increasing ζL more SMCs, which have short

75
4. Channel Characteristics and General Modeling Aspects

(a) Before: specular + diffuse (b) After: only specular

Figure 4.20: Example of applying the specular multipath extraction method.

lifetimes, are also mistakenly classified as DMCs, although a MPC that has short
lifetime could be a part of a longer but discontinuous path that is associated with
one scatterer in the propagation environment. The aforementioned effect of the ζL
parameter value on the SMCs extraction can be clearly seen in Figure 4.22.
The thresholds ζ∆τ and ζL should be chosen to minimize the SMCs to be
mistakenly classified as DMCs. However, this will also result in some DMCs to be
classified as SMCs. Nevertheless, the misclassification of some DMCs as SMCs is
affordable since their contribution to the total received power is marginal. Moreover,
most of these misclassified MPCs can later be removed based on their short lifetime
in the tracking step.

The aforementioned method is applied on the estimated MPCs in the urban

environment from Figure 4.19b. Figure 4.23a shows the CIR based on the estimated
parameters of the SMCs. The CIR of the residual after extracting the SMCs is
shown in Figure 4.23b. The values of the parameters used are summarized in
Table 4.5.

76
 4.5 MPCs Parameters Estimation and Tracking

-10 -10

Power [dB]
0.2

Power [dB]
0.2

Delay [ s]
Delay [ s]

-20 -20
-30 -30
0.1 0.1
-40 -40
0 -50 0 -50
17 17.2 17.4 17.6 17.8 18 17 17.2 17.4 17.6 17.8 18
Time [s] Time [s]

(a) ζ∆τ = 1 ns (a) ζL = 5 MPCs

-10 -10

Power [dB]
0.2

Power [dB]

Delay [ s]
0.2
Delay [ s]

-20 -20
-30 -30
0.1 0.1
-40 -40
0 -50 0 -50
17 17.2 17.4 17.6 17.8 18 17 17.2 17.4 17.6 17.8 18
Time [s] Time [s]

(b) ζ∆τ = 1.5 ns (b) ζL = 10 MPCs

-10 -10
Power [dB]

0.2

Power [dB]
0.2
Delay [ s]

Delay [ s]
-20 -20
-30 -30
0.1 0.1
-40 -40
0 -50 0 -50
17 17.2 17.4 17.6 17.8 18 17 17.2 17.4 17.6 17.8 18
Time [s] Time [s]

(c) ζ∆τ = 2 ns (c) ζL = 15 MPCs

Figure 4.21: The extracted specu- Figure 4.22: The extracted spec-
lar MPCs using different values for ular MPCs using different values
ζ∆τ while ζL = 10 MPCs and δ = for ζL while ζ∆τ =1.5 ns and δ =
25 snapshots. 25 snapshots.

(a) Specular paths extracted from the esti- (b) The residual after extracting specular
mated CIR paths. The residual may contain DMC, noise,
and artifacts.

Figure 4.23: Results of the specular MPCs extraction method.

77
 4. Channel Characteristics and General Modeling Aspects

4.5.2 Specular Multipath Tracking

Tracking the temporal evolution of the extracted SMCs is a prerequisite to localize
their scatterers using the phase changes. The KEST algorithm is able to continuously
track each estimated MPC from Figure 4.19a over time in the open-field scenario
as visualized in Figure 4.24. By visual inspection it can be seen that generally the
KEST algorithm correctly associates and tracks the MPCs over time. However, it is
not able to track the paths for a long period of time in the urban scenario due to the
richness of DMC. Therefore, in this thesis, a novel multipath component distance
(MCD)-based tracking algorithm that utilizes the estimated delay and magnitude
for tracking is proposed, and has been published in [17]. By assuming that in the
same snapshot, the delay and magnitude of each MPC are unique, only one MPC
with the same delay and magnitude exists. The main steps are as follows:

Figure 4.24: Tracked path over time in the open-field environment as a results of
the KEST algorithm.

0. Initialize the set of indices (labels) of all MPCs in the first snapshot I 1 =
{1, 2, 3, . . . , N1 }, Where N1 is the number of MPCs in the snapshot n = 1. A delay
difference threshold ζ∆τ is set to be equal to a predefined initial value ζ∆τ ini . A

78
 4.5 MPCs Parameters Estimation and Tracking

magnitude difference threshold ζ∆α is also set to be equal to a predefined initial

value ζ∆α ini .

For each MPC m = 1, 2, . . . , Nn in the snapshot n, do:

1. Find the set of indices of the MPCs from the previous snapshot that have delay
differences to τim,n less than or equal to the threshold ζ∆τ .
⃓
Am,n = {im̃,n−1 ∈ I n−1 ⃓ |τim,n − τim̃,n−1 | ≤ ζ∆τ }, (4.36)
⃓

where I n−1 is the set of all indices in the snapshot n − 1, and im̃,n−1 is the index
(label) of the MPC m̃ in snapshot n − 1.

2. Similarly, find the set of indices of the MPCs from the previous snapshot that
have magnitude differences to αim,n less than or equal to the threshold ζ∆α .
⃓
Bm,n = {im̃,n−1 ∈ I n−1 ⃓ |αim,n − αim̃,n−1 | ≤ ζ∆α }, (4.37)
⃓

for simplicity in notation, Am,n =A

ˆ︁ and Bm,n =B.
ˆ︁

3. Let
C = arg min |τim,n − τim̃,n−1 | (4.38)
im̃,n−1 ∈I n−1

be the index (label) of the MPC that has a minimum delay difference to τim,n . Only
one MPC will have a minimum delay difference.

4. Similarly, let
D = arg min |αim,n − αim̃,n−1 | (4.39)
im̃,n−1 ∈I n−1

be the index of the MPC that has a minimum magnitude difference to αim,n . Only
one MPC will have a minimum amplitude difference.

5. For each MPC m in each snapshot n, the values of both ζ∆τ and ζ∆α are updated
as follows

a. If C = D, A = ∅ and B ̸= ∅, where ∅ is an empty set, which means that

there is a MPC that has both minimum delay difference and minimum magnitude

79
 4. Channel Characteristics and General Modeling Aspects

difference. However, the delay difference is above the threshold. Therefore, the
current threshold is increased as
⎧
⎨|τi
⎪
− τn−1 (C)| if |τim,n − τn−1 (C)| < ζ∆τ ini
ζ∆τ = ⎪ (4.40)
m,n

⎩ ζ∆τ ini otherwise

where τn−1 (C) is the delay of the MPC that has an index C in snapshot n − 1.
This means that the delay difference threshold is set to be equal to the minimum
difference if this minimum does not exceed the initial delay difference threshold.
Otherwise, the threshold will be updated to be equal to the initial threshold.
The magnitude threshold is updated by a weighted average

ζ∆α = κα ∆αk (D), (4.41)

where ∆αk (D) = |αk (D) − αk−1 (D)| is the absolute of the magnitude difference
between two MPCs with the same index D in consecutive snapshots, k = n −
500, . . . , n−1 are the indices of the previous 500 snapshots, and κα > 1 is a constant
parameter. This means that the magnitude difference threshold is updated by the
weighted average of the magnitude differences in the closest tracked path.

b. If C = D, A ̸= ∅ and B = ∅, which means that there is a MPC that has

both minimum delay difference and minimum magnitude difference. However, the
magnitude difference is above the threshold. Therefore, the current magnitude
threshold is increased as
⎧
⎨|αi
⎪
− αn−1 (C)| if |αim,n − αn−1 (C)| < ζ∆α ini
ζ∆α = (4.42)
m,n

⎩ζ∆α otherwise
⎪
ini

Similar to (a), the delay threshold is updated by a weighted average

ζ∆τ = κτ ∆τk (C), (4.43)

80
 4.5 MPCs Parameters Estimation and Tracking

where ∆τk (C) = |τk (C) − τk−1 (C)| is the absolute of the delay difference between
two MPCs with the same index C in consecutive snapshots, k = n − 500, . . . , n − 1
are the indices of the previous 500 snapshots, and κτ > 1 is a constant parameter.

After updating the thresholds, the tracking begins as follows:

6. If A ∩ B ̸= ∅, one or more matching MPCs are found. If more than one matching
MPCs are found, then select the one which has the least delay difference. If the
index of the matching MPC is not already assigned to another MPC in the same
snapshot n (step 7), then Im,n = A ∩ B.

7. If A ̸= ∅, B ≠ ∅, and A ∩ B = ∅, then there are at least two different MPCs,

one of them satisfies the magnitude difference condition and the other satisfies the
delay difference condition. In order to decide which MPC to choose, a metric γ for
each choice is calculated as
(︄ )︄ (︄ )︄
ζ∆τ ζ∆α
γτ = ωτ + ωα , (4.44)
|τim,n − τn−1 (C)| |αim,n − αn−1 (C)|

and (︄ )︄ (︄ )︄
ζ∆τ ζ∆α
γα = ωτ + ωα , (4.45)
|τim,n − τn−1 (D)| |αim,n − αn−1 (D)|
where ωτ + ωα = 1, and ωτ > 0, ωα > 0 are the weight parameters of the delay and
magnitude, respectively. Based on the values of γτ and γα , the matching MPC is
then selected as ⎧
⎨C
⎪
if γτ ≥ γα
im,n = (4.46)
⎩D if γα > γτ .
⎪

Which means that matching MPC is the one that satisfies the delay condition if
γτ ≥ γα , or it is the one that satisfies the magnitude condition if γα > γτ . However,
the selection is final only if the index of the matching MPC is not already assigned
to another MPC in the same snapshot n (step 8).

8. If the index of the matching MPC is already selected by another MPC with a
delay τil,n , then the delay of both MPCs will be compared with τn−1 (C). The MPC
with the least delay difference will be assigned the index C while the other MPC

81
 4. Channel Characteristics and General Modeling Aspects

will search for a match within the previous δp snapshots (step 10), where δp is a
constant parameter corresponding to the length of the search region,
⎧
⎨C
⎪
if |τim,n − τn−1 (C)| < |τil,n − τn−1 (C)|,
im,n = (4.47)
⎩Search
⎪
back (step 10) otherwise,

and consequently,
⎧
⎨C
⎪
if |τil,n − τn−1 (C)| < |τim,n − τn−1 (C)|,
il,n = (4.48)
⎩Search
⎪
back (step 10) otherwise.

9. If A = ∅ or B = ∅, then there is no match in the previous snapshot. Therefore,

the MPC will search for a match within the previous δp snapshots (step 10).

10. Search Back: When the MPC does not find a match from the previous snapshot,
then it searches for a match within a window of length δp snapshots. For each
MPC in each snapshot in the search window, the distance metric is calculated as,
⌜ (︄ )︄2 (︄ )︄2
− τim̃,n−p αim,n − αim̃,n−p
⃓
τim,n
Sm̃,n−p = + ωα (4.49)
⃓
⎷ω ,
τ
τim,n αim,n

⃓
where S p = {Sm̃,n−p ⃓p = 2, 3, , . . . , δp + 1} is a set of all distances Sm̃,n−p , and
⃓

p = 2, 3, , . . . , δp + 1. The matching MPC is selected as follows:

⎧
⎨im̃,n−p
⎪
if Sm̃,n−p = min(S p )
im,n = (4.50)
⎩new
⎪
path otherwise,

If a matching path is found go to (step 8), and if the path is considered as a new
path, it will get a new index.

The parameters values used for tracking are summarized in Table 4.6.
The performance of the algorithm is subjected to the value of the parameters.
The parameter tuning procedure could be long and time consuming. The proposed
tracking algorithm will only require minimum offline tuning efforts. Seven parameters,

82
 4.5 MPCs Parameters Estimation and Tracking

listed in Table 4.6, need to be initialized, and two parameters, namely, the delay
difference threshold ζ∆τ and the magnitude difference threshold ζ∆α are tuned
automatically. The measured delay differences |τim,n − τim̃,n−1 | and the updated delay
threshold for the LoS path are shown in Figure 4.25a. The updated threshold is
upper-bounded by the initial threshold ζ∆τ ini =10 ns and its value for each MPC in
each snapshot is updated as in step 5. Similarly, the measured magnitude differences
|αim,n − αim̃,n−1 | and the updated magnitude threshold for the LoS path are shown
in Figure 4.25b

3
Measured magnitude changes
10
0 2.5 Updated magnitude threshold
Magnitude [dB]
Delay [ns]

10-2 1.5

1
-4 0.5
10 Measured delay changes
Updated delay threshold
0
10 12 14 16 18 20 10 12 14 16 18 20
Time [s] Time [s]
(a) Delay threshold update (b) Magnitude threshold update

Figure 4.25: The automated tuning of the delay and magnitude threshold for the
LoS component.

The initial thresholds ζ∆τ ini and ζ∆α ini represent upper bounds on the delay
and amplitude changes between two consecutive measurement snapshots. It is only
important to choose values larger than the average delay and amplitude changes
between two consecutive measurement snapshots. For example with average delay
change of 0.5 ns, any value between 5 −10 ns is suitable. Similarly, with an average
magnitude change of 0.15 dB, any value between 1.5 and 3 dB is suitable. The
constants parameter κτ is used to update the delay threshold in Step 5b. In this
initial tracking, the LoS can be easily tracked because of its high power and smooth
delay evolution over time. Based on the delay changes within the tracked LoS
path, the weighting constants are tuned to be large enough such that the updated
threshold is larger than the delay difference of the tracked LoS and small enough to
not exceed the initial threshold. Similar procedure is also applied to tune κα .

83
4. Channel Characteristics and General Modeling Aspects

When the algorithm needs to decide between two MPCs, only one satisfies the
delay condition while the other only satisfies the power condition, the weighting
constants ωτ and ωα become useful. Higher delay weight means that the delay
similarity is preferred over the power similarity. The selection of the weights values
depends on the objective of the tracking and on the data. Due to the fact that
paths with different delay may have similar power and the objective is to track the
evolution of MPCs over time, larger delay weight is selected.

Parameter Name Value

ζ∆τ ini Initial delay difference threshold 10 ns
ζ∆α ini Initial magnitude difference threshold 3.04 dB
κτ Tuning constant for delay difference threshold 10
κα Tuning constant for magnitude difference threshold 5
ωτ Delay weight 0.9
ωα Magnitude weight 0.1
δp Search range 100 snapshots

Table 4.6: Tracking parameters.

Figure 4.26 displays the results of the SMCs tracking for the urban environment
of Figure 4.19b. In the first part of the measurements until 12.5 s where Tx and Rx
were in NLoS condition, the algorithm succeeds in tracking the signal diffracted by
the obstructing building (B7). However it is unable to find singular MPCs that can
be tracked over longer periods of time inside the (B1) cloud. In the LoS situation,
when stronger and more discrete MPCs appear, the algorithm is again able to track
some strong MPCs, as for instance (B2, B3, and B5). These MPCs appear divided
in different chunks. If they belong to the same or to different objects will be assessed
in the next section. The accuracy of scatterer localization in the next section can be
considered as an indication of the performance of the proposed tracking algorithm.

84
 4.6 Scatterers Localization

Figure 4.26: Tracked path over time for Scenario 1 in the urban environment as a
results of the proposed tracking algorithm.

4.6 Scatterers Localization

The Doppler ν is estimated based on the phase of the estimated complex amplitude.
The estimated delay τ and Doppler are used to localize scatterers in the propagation
environment. For the sake of simplicity, a reflecting object is considered as a
source of single bounce reflection. A two-dimensional Cartesian coordinate system
is considered to describe the propagation scenario. The LoS and the scattered
components are assumed to propagate horizontally. This means that the differences
in antenna heights and the scattering point height are neglected and the scatterers
are then located on a ground plane together with the Tx and Rx antennas. The delay
τ of a MPC depends on the Tx position T, the Rx position R, and the scatterer
position S as
1
τ = (∥S − T∥ + ∥R − S∥) , (4.51)
c
where c is the speed of light.

85
4. Channel Characteristics and General Modeling Aspects

When only Tx or Rx is moving, the Doppler frequency ν in the 2D Cartesian

coordinate system has a shape of a hyperbola. However, if both Tx and Rx are
moving, the Doppler frequency has no typical shape any more. The Doppler frequency
is given by (︄ )︄
fc S−T R−S
ν= vT + vR , (4.52)
c ∥S − T∥ ∥R − S∥
where vT and vR are the velocity of Tx and Rx, respectively. In this representation,
the scatterer location is calculated by intersecting the curves represented by τ and ν.
The intersection results in two points on the ellipse given by Equation (4.51). Only
one of them is the true scattering point location while the other is an ambiguity. In
this thesis, the localization algorithm from [77] is employed where scatterer location
in each snapshot n and MPC m is calculated as a probability density function (PDF).
τ and ν are jointly expressed as a bivariate normal distribution, i.e., K(θ) ∼ N (µ, Σ)
in the parameter space θ
⎡ ⎤
τ
θ = ⎣ ⎦, (4.53)
ν
where ⎡ ⎤
µτ ⎦
µ=⎣ , (4.54)
µν
is the mean values, and ⎡ ⎤
σ2 0 ⎦
Σ=⎣ τ , (4.55)
0 σν2
is the covariance matrix, which is calculated from the estimated parameters in
Section 4.5 assuming that τ and ν are independent. The joint PDF of the parameter
θ is given by

1 (τ − µτ )2 (ν − µν )2
(︄ )︄
p(θ; µ, Σ) = exp − − . (4.56)
2πστ σν 2στ2 2σν2

The next step is to transform the PDF p(θ; µ, Σ) from the parameter domain into
the Cartesian domain as
1
p(X) = p(θ)|J|, (4.57)
N

86
 4.6 Scatterers Localization

where N is the number of intersections between the shape defined by τ and the
shape defined by ν, X = [x, y]T , and J is the Jacobian of θ
⎛⎡ ⎤⎞
∂τ ∂τ
J = det ⎝⎣ ∂x
∂ν
∂y
∂ν
⎦⎠ . (4.58)
∂x ∂y

The PDFs of the same MPC are then averaged over the lifetime of the MPC. During
the movement of Tx and Rx, the ambiguities change their locations while the true
locations remain fixed. As a result of the averaging, the ambiguities are partially or
completely averaged out.

Figure 4.27: Tracked paths over time in the open-field environment as a results of
KEST algorithm.

The CIR of Scenario 1 in the open-field environment is shown in Figure 4.27. Each
MPC is tracked over time and then related to a physical scatterer in the propagation
environment. Figure 4.28 shows the estimated locations of scatterers in the open-field
environment, and the photos of the scatterers are shown in Figure 4.29. The purpose
of showing the results of scatterers localization in the open-field environment is to
show the performance of the localization algorithm in an controlled environment

87
4. Channel Characteristics and General Modeling Aspects

Figure 4.28: Estimated Locations of the scatterers in open-field scenario. The results
not only show the true locations but also the ambiguities. The trajectories of the Tx
and the Rx are shown © Google.

Figure 4.29: Photos of the localized scatterers in the open-field environment.

88
 4.6 Scatterers Localization

with perfect GNSS reception and low multipath interference. The results show that
the localization algorithm is able to accurately localize the scatterers. The PDFs,
which are marked by arrows, represent the PDFs of the true location. Other PDFs
represent the ambiguities. Since only the Tx is moving, while the Rx is static, the
ambiguities do not change their position and therefore do not average out. At large
Tx-Rx separation distance, reflection occurs on the hangars (A9 and A10). As the
vehicle approaches the collision point, other scatterers become active such as the
parked vehicles in the parking lots (A1, A2, and A3), a small metal container (A4)
and a nearby metal electric box (A5). Far office buildings (A8, A11, A13, and A7)
are also identified as reflection sources. A12 and A6 are the positions of the channel
sounder and the nearby trolleys, respectively.

Figure 4.30: Tracked path over time for Scenario 1 in the urban environment as a
results of the proposed tracking algorithm.

The locations of the scatterers for all scenarios are estimated. An example of the
scatterers location is shown in Figure 4.31 for Scenario 1 in the urban environment.
In NLoS situation, several short MPCs (B1) in Figure 4.30 reach the receiver after
scattering from tree branches and leaves as shown in Figure 4.31a and Figure 4.32.

89
4. Channel Characteristics and General Modeling Aspects

(a) NLoS

(b) LoS

Figure 4.31: Estimated Locations of the scatterers in the urban Scenario 1. The
results not only show the true locations but also the ambiguities. The trajectories of
the Tx and the Rx are shown © Google. At each time instant, the trajectories of the
Tx and Rx together with the point in the center of the location PDF have the same
color.
90
 4.6 Scatterers Localization

Figure 4.32: Photos of the localized scatterers in the urban Scenario 1.

The lampposts and the traffic signs under the trees canopy (B1) could also be
sources of reflections during NLoS situation. It can be seen that the ambiguities
are partially averaged. Despite the existence of several objects, such as parked cars,
traffic signs, and the corner building, that have visibility to both the Tx and the Rx,
no MPCs are received from these objects. Simply, in NLoS scenario the location of
these objects and the position of Tx and Rx do not satisfy the law of reflection.
Single and double-bounce specular reflections from the surrounding objects in
the environment in LoS situation are observed. The estimated locations of the single
bounce reflections (B3) are not exactly on the facade of the building but rather
1 −2 m behind it as shown in Figure 4.31b. One reason could be the inaccuracy of
GNSS data. Moreover, it is assumed that the reflection point together with the
Tx and Rx antennas are located on the same horizontal plane. Therefore, when
the actual height of the reflection point does not comply with the assumption, the
estimated location will appear in front of or behind the facade of the reflectors.
In this scenario, (B3) MPCs are most probably generated by reflections from the
metallic balconies of the first floor. The estimated locations of the MPCs (B2)
appear to be about 15 m behind the building facade. It has been found, by simple
ray-tracing, that MPCs (B2) are generated by double bounce reflections from the
right then the left buildings. In addition to the single and double-bounce reflections
from the right and left building, reflections from several parked cars (B4) on the

91
4. Channel Characteristics and General Modeling Aspects

left side of the road and from a car (B6) at the corner near to the cyclist are also
observed. Moreover, the metallic shop sign is found to be the source of reflections
(B5).
The AoD and AoA of each MPC at each measurement snapshot are then
calculated geometrically from the estimated position of the last-bounce scatterer
with respect to the positions of the Tx and Rx.

4.7 Diffuse Multipath Components

When modeling the radio channel following the GSCM approach, it is often assumed
that the channel is a collection of specular and discrete multipaths that can
be associated with large and discrete objects with mirror-like reflecting surfaces.
However, [78] shows that the propagation channel does not only consist of SMCs2 ,
but also DMCs. DMC is the part of the channel energy that cannot be associated
with SMC disregarding its propagation mechanism, e.g., reflection, scattering, and
diffraction. DMC consists of large number of weak MPCs and large portion of it is
originated from diffuse scattering on rough surfaces, on small objects compared to
the signal wavelength, and on surfaces with multiple layers of different materials
[29]. DMCs could also contain SMCs that cannot be reliably detected due to their
too weak signal-to-noise-ratio (SNR) or due to specular reflection model mismatch
[79], [74]. In a propagation channel, there are relatively small number of SMC
compared to the larger number of DMC. Although the power of a single SMC
is much larger than the power of a single DMC, but due to the large number of
DMCs, their contribution to the total received power can be significant and even
dominant depending on the propagation environment. Neglecting the DMC in
channel models may result in underestimating the received power and the channel
capacity [80]. However, the contribution of the DMC is not considered in some
standard channel models such as the 3GPP-SCM [81], ITU-R M.2135 channel model
[82], and WINNER II channel model [83]. Therefore, it is of immense importance to
analyze the contribution of the DMC to the total received power in order to decide
whether or not to include the DMC in the channel model.
2
SMCs also includes the LoS component.

92
 4.7 Diffuse Multipath Components

It has been found in several studies that the DMC carries a significant fraction
of the total measured channel power. Based on indoor measurements, the authors
in [79] have found that DMC accounts for 10 %-95 % of the measured channel
power. The fractional power of the DMC is up to 70 % in an industrial environment
[84], and between 10 %-90 % in an outdoor environment [80]. In the following, the
power contribution of the DMC for the three scenarios in the urban environment is
calculated.
Having the SMCs identified and extracted in Section 4.5.1, the impulse response
that corresponds to the LoS component and the SMC, HSMC , is reconstructed based
on the estimated parameters. The impulse response of the DMC, HDMC , is then
obtained by subtracting the SMC impulse response from the measured CIR, Hmeas :

Hmeas = HSMC + HDMC . (4.59)

The total received power together with the power carried by SMC and DMC as
a function of distance are depicted in Figure 4.33 for the three scenarios. It is clear
from Figure 4.33 that the overall power is decaying with Tx-Rx distance.
In the first scenario (Figure 4.33a), the LoS is blocked by buildings when the
distance is larger than approximately 32 m. Therefore, the DMC power dominates
the SMC power and it is up to 10 dB above the power of the SMC. As the vehicle
and the cyclist approach the collision point, the SMC power increases significantly
and reaches up to 16 dB above the power of the DMC. Note that this increase
in the SMC power is due to the strong LoS and the reflections originated from
nearby buildings and parked vehicles. It has also been found that an increase in the
SMC power is often accompanied with an increase in the DMC power. This can
be explained by the fact that when a specular reflection occurs, more DMC will be
originated from the same reflecting object due to its roughness in agreement with
the findings reported in [85].
In Scenario 2 (Figure 4.33b), the cyclist is moving in the same direction as the
vehicle and the direct link between the vehicle and the cyclist repeatedly changes
between LoS and OLoS situations due to obstruction by parked vehicles. Therefore,

93
4. Channel Characteristics and General Modeling Aspects

the total received power and the SMC power experience rapid variations. The power
of the DMC is approximately up to 10 dB below SMC during the whole scenario.
Similar fluctuations are seen in Scenario 3 (Figure 4.33c). The DMC power
is also up to 10 dB below the SMC power with an exception at distance between
10 −25 m at which the power carried by the DMC is in the order of 2 −4 dB above
SMC. This can be explained by the fact that the LoS between the vehicle and the
pedestrian as well as the SMC undergo strong attenuation due to obstruction by
parked vehicles in the vicinity of the collision point.

-20 -20
Measured Measured
SMC SMC
-30 -30
DMC DMC

-40
Power [dBm]

-40
Power [dBm]

-50 -50

-60 -60

-70 -70

-80 -80
0 20 40 60 80 100 0 20 40 60 80 100
Tx-Rx distance [m] Tx-Rx distance [m]
(a) Scenario 1 (b) Scenario 2

-20
Measured
SMC
-30
DMC

-40
Power [dBm]

-50

-60

-70

-80
0 20 40 60 80 100
Tx-Rx distance [m]
(c) Scenario 3

Figure 4.33: Total received power and power carried by the SMC and DMC as a
function of Tx-Rx distance for the three scenarios in the urban environment.

94
 4.7 Diffuse Multipath Components

An important aspect of modeling the DMC is the fractional DMC power f̂ DMC ,
i.e., the percentage of the DMC power contribution in the total received power:

PDMC
f̂ DMC = × 100 %, (4.60)
PDMC + PSMC

where PDMC and PSMC are the average power carried by DMC and SMC respectively.
Figure 4.34 presents the fractional DMC power f̂ DMC calculated with Equation
(4.60) for the three scenarios. It can be seen in Figure 4.34 that the fractional DMC
power in the first scenario exceeds 40 % and reaches up to 90 % of the total received
power in NLoS situation at distances larger than approximately 32 m. As the vehicle
and the cyclist are moving toward the intersection, the contribution of the DMC to
the total received power decreases to reach approximately 5 % at the collision point.
In the second scenario, the fractional DMC power varies between 15 %-40 %. The
variations are aligned with the variation of the SMC power in Figure 4.33b. The
contribution of the DMC decreases to around 5 % at the collision point due to the
strong LoS component, which usually carries most of the transmitted power.
Similar to the second scenario, the fractional DMC power in the third scenario
represents from 15 % to 40 % of the total received power at distances larger than
25 m. A rapid increase to around 60 % shortly followed by a sudden drop to around
10 % prior to the collision is observed. Note that these results are in agreement with
the results reported in [80].
From the results it becomes clear that the amount of the fractional DMC power
depends on the link shadowing condition. Therefore, it is larger in NLoS than
in OLoS and LoS situations. An increase of the fractional power can be mainly
attributed to the decreased power of the LoS component due to obstruction by,
e.g., buildings, parked and moving vehicles, etc. Similarly, the smaller amount of
DMC power in LoS situation is probably due to the existence of a dominant LoS
component and strong specular reflections.

95
4. Channel Characteristics and General Modeling Aspects

100
Scenario 1
Scenario 2
Contribution of DMC [%]

80 Scenario 3

60

40

20

0
0 20 40 60 80 100
Tx-Rx distance [m]
Figure 4.34: The fractional DMC power f̂ DMC as a function of Tx-Rx distance
calculated by Equation (4.60) for the three scenarios in the urban environment.

4.8 Summary
This chapter addresses several aspects of channel modeling based on measurements
in both open-field and urban environments. The presented results and methods in
this chapter have been published in [16, 17, 22, 23, 24, 25].
Path loss models for the measured scenarios in open-field environments are
proposed. The two-ray model is found to provide a good fit to the measured path
loss in LoS scenario with static tripod. In Scenario 2, rapid fluctuations of the
measured path loss are observed. These fluctuations are the result of the MPC
originated by the human body. Further, the moving human body also results in
dynamic antenna height and position. Due to the shadowing effect caused by the
crowd around the Rx in Scenario 3, the measured path loss value is 5 −10 dB larger
than the LoS case. In Scenario 4, the obstruction of the LoS by parked vehicles
causes an extra loss between 10 − 20 dB depending of the size of the parked vehicle.
Log-distance path loss models are proposed for Scenario 2, Scenario 3, and Scenario
4. Further, the spatial correlation of shadow fading for Scenario 1 and Scenario

96
 4.8 Summary

4 of the open-field environment is investigated. First, to prove the stationarity

assumption, the SF is put under test, namely, the MRA test. After passing the
stationarity test, the autocorrelation functions of the shadow fading are estimated
and modeled by three different theoretical autocorrelation function models. The
results show that the exponential decaying sinusoid model provides the best match
to the empirical autocorrelation function in all scenarios. To study and model the
aforementioned loss in the received power due to obstruction of LoS by parked
vehicles in Scenario 4, a 3D ray tracing tool is developed. The measured path loss
within the obstruction region is compared with the modeled path loss using the
combined two-ray path loss model and the multiple knife-edge diffraction model.
Despite its simplicity, the proposed model is able to provide a good match for the
measured path loss.
From the measurements in the urban environment, single and dual-slope log-
distance path loss models are proposed. From the results, it is noticed that due to
the obstruction by buildings in Scenario 1, an extra loss of up to 25 dB is experienced
when compared to the free space path loss. Moreover, due to obstruction of parked
vehicles, a diffraction loss between 5 dB and 15 dB is observed in Scenario 2, and
Scenario 3, respectively.
The non-stationarity of the V2VRU channel in the urban environment is assisted
by estimating the generalized local scattering function and its collinearity based on
the channel measurement data. It is found that in all scenarios, at the shortest Tx-Rx
distance, i.e., at the collision point, the stationarity distance suddenly drops. This
results from the sudden change in the Doppler domain because the LoS component
shifts from positive to negative Doppler at the moment of passing the collision point.
Furthermore, a pattern of consecutive highs and lows is observed, which can be
explained by the appearance and disappearance of the LoS due to the blockage by the
parked vehicles. Generally, a stronger LoS component leads to a large stationarity
distance. Moreover, not only the power of the LoS but also its variation affects
the stationarity distance. It is noticed that the that larger and more rapid the
power variations of the LoS component the shorter the stationarity distance. The
estimated average stationarity distances are 2.13 m, 5.25 m, and 3.58 m for Scenario
1, Scenario 2, and Scenario 3, respectively.

97
4. Channel Characteristics and General Modeling Aspects

The KEST algorithm is employed to estimate the parameters of the MPCs.

Unlike the open-field, in the urban measurements, the CIR is highly cluttered. It
can be deduced based on the long lifetimes that all paths in the estimated CIR
in the open-field are due to specular reflections. On the other hand, in the urban
scenario, not only specular reflections but also DMCs can be seen. Therefore, the
SMCs are extracted from the estimated time-variant CIR. For this purpose, a novel
method to separate SMCs and DMCs based on their number of neighboring MPCs
that have delay difference less or equal a specific threshold is proposed. Next, an
MCD-based tracking algorithm that utilizes the estimated delay and magnitude for
tracking the SMCs is proposed. Having the SMCs tracked over time, each SMC is
related to a physical scatterer in the propagation environment. Localization of the
scatterers in all scenarios is done using a joint delay-Doppler estimation algorithm.
The estimated positions of the scatterers are then used to estimate the AoD and
AoA.
Finally, the contribution of the DMC to the total received power is calculated.
From the results it becomes clear that the amount of the fractional DMC power
depends on the link shadowing condition. Therefore, it is larger in NLoS than
in OLoS and LoS situations. An increase of the fractional power can be mainly
attributed to the decreased power of the LoS component due to obstruction by,
e.g., buildings, parked and moving vehicles, etc. Similarly, the smaller amount of
DMC power in LoS situation is probably due to the existence of a dominant LoS
component and strong specular reflections. It is found that the DMC carries up to
5 % of the total received power in LoS condition. However, in OLoS the fractional
DMC power is approximately between 15 % and 40 % and it can reach 90 % in NLoS
condition.
Having the SMCs extracted, tracked over time, and their scatterers localized,
a full parameterization of the GSCM channel model becomes feasible. The next
chapter will address the estimation of the model parameters based on the estimated
amplitude, delay, and angles of the SMCs.

98
 Channel Model Parameterization
5
For developing a GSCM for V2VRU communication following WINNER II channel
model, the LSPs need to be estimated. In this chapter, the LSPs, introduced in
Section 2.4.1, are estimated in the power and delay domain, i.e., SF, DS, and KF,
and in the angular domain, i.e., ASD, and ASA. In order to maintain the spatial
correlation of the LSPs observed in the measured channel, the autocorrelation of
these LSPs is analyzed and the correlation distances are calculated. Furthermore,
to ensure the spatial consistency, the cross-correlation coefficients among the LSPs
are calculated. The model parameters are then used as an input to the channel
simulator.

5.1 Large Scale Parameters

In this section, the estimation of the first- and second-order statistics of the LSPs
are provided. The LSPs are estimated from the specular MPC parameters, i.e., the
power, delay, and angles, that were previously estimated from the measurement data
in Section 4.5 and from scatterer locations in Section 4.6. The LSPs are then fitted
to log-normal distributions. These distributions control the behavior of the modeled
channel. Table 5.1 summarizes the proposed LSPs of the channel model.

99
5. Channel Model Parameterization

Due to the non-stationarity of the channel, the channel is divided into regions,
within which the WSS assumption holds. The LSPs and their correlations are
evaluated within these regions using a sliding average window of length of 15 λ
(0.86 m). The length of the window, i.e., the stationarity distance, is found to be
valid for all scenarios based on the non-stationarity analysis in Section 4.4 (see
Figure 4.15). Note that, this sliding averaging window is different from the sliding
average window of length of 10 λ used for removing the small scale fading in Sections
4.1 and 5.1.1.
The KF and DS are found to vary significantly when the propagation condition,
i.e, LoS, OLoS, and NLoS, changes. Therefore, when estimating the KF and DS,
the data set is divided into two parts based on the propagation condition. The
appearance of the LoS starts at an average Tx-Rx distance of 15 m in Scenario 1,
and at 9 m in both Scenario 2 and Scenario 3. Unlike the KF and DS, the SF, ASD,
and ASA are estimated for the whole experiment with no distinction between LoS
and OLoS/NLoS situations.

5.1.1 Shadow Fading

As explained in Section 4.1, the measured path loss is expressed in logarithmic
scale as a sum of two components, namely, the the distance-dependent path loss
model and the SF (see Equation (4.5)). To remove the effect of fast fading from
the measured path loss, a sliding window of length 10 λ is used. The SF is usually
modeled as a normally-distributed random variable with zero mean and a standard
deviation. In all three scenarios, the SF is estimated for the whole experiment, i.e.,
the data set is not divided based on the propagation condition.
Figure 5.1 shows the results for each scenario. Each plot depicts the CDF curves
of the measured SF and its log-normal fit. Positive SF values indicate a larger overall
path loss, while negative SF values mean a smaller overall path loss. Even though
minor differences can be noticed in Scenario 3 between the CDF of the measured SF
and its log-normal fit, it can be seen that the log-normal distribution perfectly fits
the distribution of the measured SF with a zero mean and a standard deviation σ.

100
 5.1 Large Scale Parameters

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 5.1: CDFs of measured SF and the log-normal fit for the three scenarios in
urban environment.

101
5. Channel Model Parameterization

5.1.2 Delay Spread

The DS is widely used to statistically describe delay characteristics of wireless
channels. Due to the multipath propagation, the Rx receives several copies of the
transmitted signal. Each copy arrives from different direction with different gain and
delay. The delay dispersion of the propagation paths is described by the DS metric.
The DS in the snapshot at time t is calculated from the estimated parameters of
the SMCs described in Section 4.5 according to [86]
⌜
⃓
⃓ N∑︂
(t)−1
Pi τi2
⃓
⃓
⃓
στ = (5.1)
⃓ i=0
⃓
⃓ N∑︂(t)−1
− τ̄ 2 ,
⃓
⎷ Pi
i=0

where τ̄ is the mean excess delay calculated as in [86]

N (t)−1
∑︂
Pi τi
τ̄ = i=0
N (t)−1
, (5.2)
∑︂
Pi
i=0

and where Pi and τi are the power and delay of the i-th MPC, and N (t) is the number
of MPCs at time t. The DS is calculated for each scenario from all measurement
runs and then fitted to a log-normal distribution with a unit of log10 (s).
Figure 5.2 shows the CDF of the measured DS alongside the log-normal fit for the
three scenarios. Each scenario is divided into two parts according to the propagation
condition. Therefore, each plot includes CDFs for two propagation conditions. In
Scenario 1, the DS has a mean value of 36 ns in NLoS case. As the vehicle and the
cyclist are approaching the intersection, the mean value drops to 7.4 ns. In Scenario
2, the mean value of the DS decreases from 25 ns to 6.4 ns when moving from OLoS
case to LoS case. Similarly, in Scenario 3, the mean values are 24 ns and 8.1 ns in
OLoS and LoS cases, respectively. Generally, the DS tends to decrease when the
power of LoS path increases. It can be observed, in all scenarios, that the CDF of

102
 5.1 Large Scale Parameters

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 5.2: CDFs of measured DS and the log-normal fit for the three scenarios in
urban environment.

103
5. Channel Model Parameterization

the log-normal fit in OLoS and NLoS has a tail with larger DS than that of the
measured one. Nevertheless, the log-normal distribution provides a good fit to the
measured DS.

5.1.3 Narrowband K-factor

The KF is an important parameter that provides an indication of the energy
proportion of the MPCs and hence quantifies the multipath richness of the
communication channel. The KF is defined as the ratio of the energy of the
dominant component that consists of the LoS component and the unresolved ground
reflection to the energy of all other components. In literature, several approaches
are proposed to estimate the KF. A widely used technique to estimate the KF
from the measured CIR is introduced in [87] based on the method of moments
(MoM). However, as stated earlier, the narrowband KF in this thesis is calculated
for each measurement snapshot from the estimated parameters of the SMCs and it
is calculated according to [86]

PLoS
K= , (5.3)
P − PLoS
where PLoS is the power of the LoS path, and P is the total power. Typically the
KF is in dB and it is modeled by a log-normal distribution.
Figure 5.3 shows the CDFs of the measured KF and its log-normal fit for the
three scenarios. Note that, the KF is not estimated for NLoS condition in Scenario
1. It can be seen that the log-normal distribution provides a perfect fit for the
measured KF. The KF in LoS case for the three scenarios has comparable mean
values with 15.02 dB, 16.22 dB, and 17.32 dB for Scenarios 1, 2, and 3, respectively.
Similarly, in OLoS case the mean values of the KF are 5.34 dB and 6.44 dB for
Scenario 2 and Scenario 3, respectively. The main reason for the larger KF in during
the LoS compared to the OLoS case is the the existence of a unobstructed strong
LoS path. In OLoS case, the LoS path is obstructed by parked vehicles and therefore
suffers from an attenuation due to the diffraction loss, as explained in Section 4.3.

104
 5.1 Large Scale Parameters

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 5.3: CDFs of measured KF and the log-normal fit for the three scenarios in
urban environment.

105
5. Channel Model Parameterization

5.1.4 Angular Spread

The AS is the parameter that describes the dispersion of propagation paths in the
angular domain, i.e., it measures how the multipath power is spread out in the
spatial domain. The AS is calculated utilizing the paths angles, i.e., AoD and AoA,
and the path gains. The AoA and AoD are calculated from the scatterers locations,
which were estimated in Section 4.6. The reference angle, i.e., the angle 0◦ is set to
be the angle between the Tx and the Rx. The ASD and ASA are calculated in each
measurement snapshot as in [88]
⌜
⃓
⃓ N∑︂
(t)−1 (︂ (︂ )︂)︂2
ξ ϕi − ϕ̄ Pi
⃓
⃓
⃓
σϕ = (5.4)
⃓ i=0
⃓
⃓ N (t)−1
,
⃓ ∑︂
⎷ Pi
i=0

where Pi and ϕi are the power and angle of the i-th MPC, function ξ (.) maps
angles to (-180, 180) degree, and the power weighted mean angle ϕ̄ is given by

N (t)−1
ξ (ϕi ) Pi
∑︂

ϕ̄ = i=0
N (t)−1
. (5.5)
∑︂
Pi
i=0

The ASD and the ASA are estimated for the whole experiment because, unlike
the DS and KF, the difference in angle spread in different propagation cases, i.e.,
LoS, OLoS, and NLoS is found to be insignificant. Figure 5.4 presents the CDFs of
the measured ASD together with its log-normal fit. The measured ASD follows the
log-normal fit in all three scenarios with minor mismatch. The mean ASD values
are 5.49◦ , 7.24◦ , and 10.96◦ for Scenarios 1, 2, and 3, respectively. The CDFs of the
ASA are plotted in Figure 5.5 which also perfectly follow the log-normal distribution
with mean values of 10.47◦ , 10.72◦ , and 12.88◦ for Scenarios 1, 2, and 3, respectively.

106
 5.1 Large Scale Parameters

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 5.4: CDFs of measured ASD and the log-normal fit for the three scenarios in
urban environment.

107
5. Channel Model Parameterization

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 5.5: CDFs of measured ASA and the log-normal fit for the three scenarios in
urban environment.

108
 5.2 Correlation Distances

5.2 Correlation Distances

The correlation distance of a LSP indicates how long the channel can be assumed
stationary for this specific LSP. The correlation distance is used as a model parameter
to maintain the spatial correlation of the LSP observed in the measured channel.
It is employed in the generation the correlated 2D map of the LSP by determining
the length of the channel segment, or the length of the traveled distance, in these
maps as discussed in Section 2.4.2. The correlation distance is calculated from the
autocorrelation function of the evaluated LSP as follows

r(∆d) = E [x(d)x(d + ∆d)] , (5.6)

where E [.] denotes the expected value, d is the distance between Tx and Rx,
∆d is the change of the Tx-Rx distance, and x is the LSP under evaluation in log
domain. The correlation distance dc is defined as the value of ∆d at which the
value of the normalized autocorrelation function drops to e−1 . Note that, in case
the normalized autocorrelation function has a value of e−1 at different values of ∆d,
the shortest ∆d is the correlation distance (see Figure 5.6). Table 5.1 summarizes
the correlation distances of the LSPs in all scenarios.
The results show that the correlation distances of the KF and DS during LoS
part are much shorter than in NLoS/OLoS part. This is attributed to the fact that
the length of the LoS part, within which the autocorrelation function is evaluated, is
limited to only few meters prior to the collision point, i.e., less than 10 m in Scenario
1 and less than 5 m in Scenario 2 and Scenario 3. Figure 5.6 shows an example of
the normalized autocorrelation function of the KF in Scenario 3 for both LoS and
OLoS parts. It can be noticed that the autocorrelation function within this distance
approximately matches the autocorrelation function during OLoS part. Similar
agreement is also found for DS in all scenarios. The findings show that the smaller
correlation distance in LoS part is only due to the shorter distance within which the
autocorrelation function is evaluated and, hence, it does not imply less correlation.
As can be seen from Table 5.1, on the one hand, the KF, DS, ASD, and ASA
show relatively high spatial correlation which implies slow change in their values

109
5. Channel Model Parameterization

1
LoS part
OLoS part
0.8
rx( (d))/rx(0)

0.6

0.4

0.2

0
0 20 40 60 80
d [m]

Figure 5.6: Example of the normalized autocorrelation function of the KF in Scenario

3 for both LoS and OLoS parts. The black dashed line marks the value of e−1 at
which the correlation distance is calculated.

over time. On the other hand, the SF experiences less spatial correlation which can
be explained by the intermittent blockage of the LoS by parked vehicles.

110
 5.2 Correlation Distances

Scenario 1 Scenario 2 Scenario 3

Parameters
LoS NLoS LoS OLoS LoS OLoS
µ 0 0 0 0 0 0
SF [dB]
σ 3.90 3.90 1.80 1.80 4.69 4.69
[Link] [m] dc 8.20 8.20 2.27 2.27 7.60 7.60
µ 15.02 N/A 16.22 5.34 17.32 6.44
K-factor [dB]
σ 4.81 N/A 2.88 4.38 4.09 5.31
[Link] [m] dc 8.01 N/A 4.42 30.58 4.23 26.97
µ -8.13 -7.44 -8.19 -7.60 -8.09 -7.62
DS [log10 s]
σ 0.34 0.16 0.19 0.21 0.23 0.22
[Link] [m] dc 8.97 49.69 4.42 70 4.64 57.88
µ 0.74 0.74 0.86 0.86 1.04 1.04
ASD [log10 ◦ ]
σ 0.50 0.50 0.39 0.39 0.30 0.30
[Link] [m] dc 37.11 37.11 63.23 63.23 55.61 55.61
µ 1.02 1.02 1.03 1.03 1.11 1.11
ASA [log10 ◦ ]
σ 0.4 0.4 0.39 0.39 0.33 0.33
[Link] [m] dc 48.95 48.95 66.17 66.17 54.48 54.48

Table 5.1: Large scale parameter distributions and correlation distances.

111
5. Channel Model Parameterization

5.3 Cross-correlation Parameters

In order to ensure channel spatial consistency, the cross-correlations between all
LSP pairs are needed. This metric describes the linear dependency between the
LSPs. The cross-correlation coefficient is given by

M
(x(m) − x̄) (y(m) − ȳ)
∑︁

ρx,y = √︄ m=1
, (5.7)
M M
(x(m) − x̄) 2
(y(m) − ȳ)
2
∑︁ ∑︁
m=1 m=1

where x and y are the measured sequences of the two LSPs pairs with length M ,
and x̄ and ȳ are the sample means of x and y, respectively.
Cross-correlation coefficients take values from -1 to 1. A negative coefficient
indicates a decrease in parameter x if parameter y increases and vice versa. Positive
coefficient means that x and y increase and decrease simultaneously. When the
coefficient is zero there is no dependency between the two parameters. Table 5.2
shows the cross-correlation coefficients between all pairs of LSPs for all scenarios in
the urban environment.
The SF is found to be positively correlated with the KF while negatively correlated
with the DS, i.e., when the SF increase, the K-factor increases and the DS decreases.
In all scenarios, the DS and KF have a strong negative correlation as expected. A
positive correlation in all scenarios between the ASD and ASA is also observed.
The ASD cross-correlations with SF, KF, and DS in the three scenarios are not
consistent. For example, the ASD has positive correlation of 0.28 with the SF in
Scenario 1, negative correlation of -0.21 in Scenario 2, and no correlation in Scenario
3. Nevertheless, the ASD shows weak or no dependency on the SF, KF and DS.
Similarly, the ASA shows no correlation with the KF and DS in all scenarios, while
it has a weak dependency on the SF in Scenarios 2 and 3.

112
 5.3 Cross-correlation Parameters

Cross-correlation SF KF DS ASD ASA

SF 1 0.5 -0.42 0.28 -0.07
KF 1 -0.86 0 -0.01
Scenario 1 DS 1 -0.33 -0.04
ASD 1 0.58
ASA 1
SF 1 0.35 -0.3 -0.21 -0.36
KF 1 -0.87 0.25 -0.01
Scenario 2 DS 1 -0.13 0.07
ASD 1 0.63
ASA 1
SF 1 0.63 -0.51 -0.07 -0.26
KF 1 -0.85 0 -0.1
Scenario 3 DS 1 0.02 -0.07
ASD 1 0.5
ASA 1

Table 5.2: Estimated cross-correlation coefficients.

113
5. Channel Model Parameterization

5.4 Scaling Coefficients and Number of MPCs

As stated in Section 2.4.1 and Section 2.4.2 (Step C.), the scaling coefficients, also
called the proportionality factors, are used to scale path delays and angles to insure
that the differences in the spreads are reflected in the powers. The scaling coefficient
of the delay g DS is defined as the ratio between the standard deviation of the path
delays and the RMS DS. Similarly, the scaling coefficients of the angles, g ASD and
g ASA , are defined as the ratio between the standard deviation of the path angles and
the AS [15]. Table 5.3 summarizes the mean values of scaling coefficients estimated
from the measurement data in all scenarios.

Scaling coefficient Scenario 1 Scenario 2 Scenario 3

g DS 1.97 2.71 3.48

g ASD 2.57 2.35 1.82
g ASA 1.98 2.38 2.18

Table 5.3: Estimated scaling factors in all scenarios.

Based on the measurement data, the number of MPCs used in the channel
simulations for all scenarios is set to 8 MPCs. However, we noticed that changing
the number of MPCs has an insignificant impact on the simulated channels.

5.5 Summary
In this chapter, a full parametrization for the WINNER-type GSCM is proposed.
The LSPs are estimated in log domain and fitted to the log-normal distribution. The
results show that the log-normal distribution provides a good fit to the measured
LSPs in all scenarios. Furthermore, the spatial correlation of the LSPs are evaluated
and their correlation distances are calculated. It can be noted that the SF experiences
relatively low correlation. However, all other LSPs show high correlation indicating
slow change in their statistics over time. Moreover, to include the inter-dependency
between all pairs of LSPs, the cross-correlation coefficients are provided.

114
 5.5 Summary

The parameters estimated in this chapter together with the path loss models are
used to generate synthetic channels. The proposed channel model is then validated
by comparing the measured channels with the simulated synthetic channels in the
next chapter.

115
 Model Validation
6
This chapter presents the validation of the proposed channel model. The GSCM for
V2VRU communications is validated by comparing the simulated channels with the
measured channels. The simulated channels are generated by the WINNER-type
QuaDRiGa simulator [40] described in Section 2.4.2. The model parameters are
extracted from the measured channels and used as an input to the simulator. These
parameters are, mainly, the path loss presented in Chapter 4, and the LSPs estimated
in Chapter 5. The channel model is validated by comparing the distributions of the
model parameters, extracted from the simulated channels, with their counterparts
extracted from the measured channels and used as input to the proposed model.
The model parameters extracted from the measurement are referred to as the model
input. In addition to the distributions of the LSPs, the correlation distance of each
LSP as well as the cross-correlation between each LSPs pair are also considered in
the validation process.
The validation is performed using a group of 100 simulated channels produced by
100 simulation runs for each scenario. These simulation runs are carried out using
Tx and Rx routes similar to the routes during the channel measurements. Therefore,
the resulting simulated channel in each run has a similar number of snapshots as the
measured channel. Furthermore, the channels are generated for a center frequency

117
6. Model Validation

Figure 6.1: Channel model validation process.

of 5.2 GHz and a bandwidth of 120 MHz. It must be also noted that only specular
MPCs are taken into account when estimating the model parameters and generating
the simulated channels. The diffuse multipath are not taken into account in this
model. Figure 6.1 shows the procedure for the channel model validation.

6.1 Simulated Channels and Path Loss

Figure 6.2 shows examples of measured and simulated CIRs for the three scenarios in
urban environment. In all scenarios the Tx-Rx separation distance is approximately
100 m. In Scenario 1 (Figure 6.2a and 6.2b), the vehicle and the cyclist are
approaching the collision point at the intersection while the LoS is completely

118
 6.1 Simulated Channels and Path Loss

obstructed by buildings. It can be seen that the LoS path starts to appear at
approximately 23 m or 6.5 s in the measured CIR and at approximately 15 m or 7 s
in the simulated channel. This difference is due to setting the Tx-Rx distance of the
LoS appearance in the simulations to 15 m, which is the average distance extracted
from all measurement runs in Scenario 1. As shown in both measured and simulated
CIRs, the LoS path is accompanied by a group of strong MPCs. During NLoS, a
weak path appears as a tail of the LoS path with a slightly larger delay than the
geometric LoS (GLoS). A group of weak paths can also be recognized.
In Scenario 2 (Figure 6.2c and 6.2d), the cyclist is approaching the collision
point while driving parallel to the vehicle. The LoS in both measured and simulated
channels is partially obstructed by parked vehicles except at distances less than
9 m where a strong LoS can be noticed. In contrary to the LoS path, which has
a propagation delay calculated in a deterministic manner from the positions of
the Tx and Rx, the propagation delays of the SMCs depend on the locations of
the scatterers which are randomly placed based on the the distributions of the AS.
Therefore, we can not expect to generate simulated channels with SMCs identical to
the measured ones.
In Scenario 3 (Figure 6.2e and 6.2f), the pedestrian is crossing the street while
the vehicle is approaching the collision point. Similar to Scenario 2, the LoS path is
partially obstructed by parked vehicles. A strong LoS appears prior to the collision
accompanied with strong MPCs generated by scatterers located in the vicinity of
the collision point. In all scenarios, it can be noticed that the SMCs during the LoS
condition are more closely located nearby the LoS path than in the OLoS and NLoS
conditions.
In order to validate that the proposed channel model produces accurate path loss
or received power, the local mean path loss, calculated from all simulated channels, is
compared with the path loss model presented in Figure 4.2 for each scenario. These
proposed log-distance path loss models are used as an input to the channel simulator.
As can be seen in Figure 6.3, in Scenario 1, the measured path loss is modeled
by a single slope in LoS propagation condition with path loss exponent of 1.4 for
distances up to 15 m from the collision point. At longer distances, up to 100 m, the
propagation condition is NLoS and the path loss is modeled by a dual-slope path

119
6. Model Validation

0.8 -10 0.8 -10

0.7 -15 0.7 -15

0.6 -20 0.6 -20

Power [dB]

Power [dB]
Delay [ s]

0.5 -25 0.5 -25

Delay [ s]
0.4 -30 0.4 -30

0.3 -35 0.3 -35

0.2 -40 0.2 -40

0.1 -45 0.1 -45

0 -50 0 -50
0 2 4 6 8 9 0 2 4 6 8 9
Time [s] Time [s]
(a) Scenario 1 - measured (b) Scenario 1 - simulated

0.8 -10 0.8 -10

0.7 -15 0.7 -15

0.6 -20 0.6 -20

Power [dB]

Power [dB]
0.5 -25 0.5 -25
Delay [ s]
Delay [ s]

0.4 -30 0.4 -30

0.3 -35 0.3 -35

0.2 -40 0.2 -40

0.1 -45 0.1 -45

0 -50 0 -50
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Time [s] Time [s]
(c) Scenario 2 - measured (d) Scenario 2 - simulated

0.8 -10 0.8 -10

0.7 -15 0.7 -15

0.6 -20 0.6 -20

Power [dB]

Power [dB]
Delay [ s]

0.5 -25 0.5 -25

Delay [ s]

0.4 -30 0.4 -30

0.3 -35 0.3 -35

0.2 -40 0.2 -40

0.1 -45 0.1 -45

0 -50 0 -50
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time [s] Time [s]
(e) Scenario 3 - measured (f) Scenario 3 - simulated

Figure 6.2: Examples of measured and simulated channel impulse responses for the
three scenarios in urban environment.
120
 6.1 Simulated Channels and Path Loss

110 110

100 100
Path loss [dB]

Path loss [dB]

90 90

80 80

70 70
Model input NLoS slope 1
Model input NLoS slope 2 Model input OLoS
60 60 Model input LoS
Model input LoS
Simulated channel Simulated channel
50 50
5 10 20 30 50 70 100 5 10 20 30 50 70 100
Tx-Rx distance [m] Tx-Rx distance [m]
(a) Scenario 1 (b) Scenario 2

110

100
Path loss [dB]

90

80

70
Model input OLoS
60 Model input LoS
Simulated channel
50
5 10 20 30 50 70 100
Tx-Rx distance [m]
(c) Scenario 3

Figure 6.3: Comparison of the path loss for the three scenarios in urban environment.

121
6. Model Validation

loss with exponents of 6.9 and 2.9. In Scenario 2, path loss exponents of 1.7 and
2.4 were obtained in LoS and OLoS propagation conditions, respectively. Similar
to Scenario 2, the path loss exponents in Scenario 3 are found to be 1.8 and 3.2 in
LoS and OLoS conditions. For more details on the proposed model parameters (see
Table 4.2).
Figure 6.3 depicts the results for the path loss validation in the three measured
scenarios. Each plot includes two or three solid colored lines and a black dashed
curve. The solid colored lines represent the model input of the path loss extracted
from the measured channels with Tx-Rx distance (see Figure 4.2), while the dashed
black curve is the local mean of the path loss obtained from the 100 simulated
channels. It can be noted that the channel model clearly provides an almost perfect
match in terms of path loss and received power in all three measured scenarios.

6.2 Large Scale Parameters

In this section, the distributions of the LSPs from the model input and the simulated
channels are compared. The LSPs are estimated from the simulated channels in
the same way they were estimated from the measured channels, as presented in
Chapter 5, and then fitted to log-normal distributions. The mean and standard
deviation values of the LSPs are summarized in Tables 6.1, 6.2, and 6.3.

6.2.1 Shadow Fading

Figure 6.4 depicts the results for the SF validation in the three measured scenarios.
Each plot includes two CDF curves. The solid orange line represents the model
input of the SF extracted from the measured channels, while the dotted black line
shows the SF obtained from the simulated channels.
It can be noted that the match between the simulations and the model input
is good for Scenario 1. The standard deviation values of the SF are 3.9 dB and
3.35 dB for the model input and simulated channels, respectively. In Scenario 2,
some mismatch can be seen. The estimated standard deviation from the simulated
channels is 1 dB larger than the SF of the model input. Similar mismatch is observed

122
 6.2 Large Scale Parameters

in Scenario 3. The simulated channels produce larger SF with a standard deviation

of 6.31 dB compared to 4.69 dB in the model input. It has been found that the
observed mismatch occurs due the final scaling of the path powers by the KF in
Equation (2.26). In Scenario 1, the impact of the KF on the SF is less than in

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 6.4: Comparison of the distribution of the shadow fading for the three
scenarios in urban environment.

Scenario 2 and Scenario 3 because the path powers are only scaled by the KF in the
channel segments with a LoS condition. However, in Scenario 2 and Scenario 3 the
path powers are scaled by the KF in all channel segments which may explain the
larger mismatch.

123
6. Model Validation

6.2.2 Delay Spread

The RMS DS is widely used to statistically describe delay characteristics of wireless
channels. The DS is modeled as a log-normal distribution with unit of log10 (s). It
is evaluated with distinction between LoS, OLoS, and NLoS propagation conditions.

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 6.5: Comparison of the distribution of the RMS delay spread for the three
scenarios in urban environment.

The LoS only corresponds to the part of the channel prior to the collision in
which the Tx-Rx distance is shorter than 15 m in Scenario 1 and 9 m in Scenario 2
and Scenario 3. The results from the DS are presented in Figure 6.5. As expected,

124
 6.2 Large Scale Parameters

the DS in LoS condition is smaller than the DS in NLoS condition in all scenarios.
It can be seen that the results from the simulated channels and the model input
have almost perfect match in all scenarios in both the LoS and OLoS propagation
conditions. In Scenario 1, the measured mean values of the DS are 7.41 ns and
36.31 ns for the LoS and NLoS, respectively. The DS mean values from the simulated
channels follow the measured ones with 8.32 ns and 34.67 ns for the LoS and NLoS,
respectively. Similar agreement is also observed between the model input and the
simulated channels in terms of the mean and the standard deviation values of the
DS in all scenarios as summarized in Tables 6.1-6.3.

6.2.3 Narrowband K-factor

The KF provides an indication of the energy proportion of the MPCs and hence
quantifies the multipath richness of the communication channel. The KF is defined
in Equation (5.3) as the ratio of the energy of the dominant component that consists
of the LoS component and the unresolved ground reflection to the energy of all other
components. Similar to the DS, the KF is also validated with distinction between
LoS, OLoS, and NLoS propagation conditions, as shown in Figure 6.6. The KF is
calculated only for the LoS part of the channel in Scenario 1. In Scenario 2 and
Scenario 3, the KF is calculated for both the LoS and the OLoS parts of the channel.
In Scenario 1, the mean and standard deviation values of the KF in the model input
are 15.02 dB and 4.81 dB, respectively. The simulated channels show almost perfect
agreement in terms of the KF with mean of 14.44 dB and standard deviation of 4.98
dB.

125
6. Model Validation

It can also be observed that the KF values extracted from the simulated channels
in Scenario 2 and Scenario 3 almost perfectly follow the KF in the model input in
both LoS and OLoS propagation conditions.

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 6.6: Comparison of the distribution of the narrowband K-factor for the three
scenarios in urban environment.

6.2.4 Angular Spread

In contrary to the KF and DS, the angular spreads are evaluated with no distinction
between LoS, OLoS, and NLoS propagation conditions. The complete data set is

126
 6.2 Large Scale Parameters

used to estimate the ASD and ASA. Figure 6.7 depicts the comparison between the
simulated channels and the model input in terms of the ASD. It can be observed
that the CDFs from the model input have tails with significantly larger ASD than
those of the simulated channels.

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 6.7: Comparison of the distribution of the azimuth spread of departure for
the three scenarios in urban environment.

The simulated channels in Scenario 1 result in larger ASD with mean of 7.76◦ in
comparison to 5.50◦ in the model input. However, better agreement can be seen in
Scenario 2 and Scenario 3. The mismatch in Scenario 1 could be attributed to the
errors in estimating path angles in NLoS. As discussed in Chapter 4, the locations of

127
6. Model Validation

scatterers are estimated using the delay and Doppler of the MPCs. These locations

(a) Scenario 1 (b) Scenario 2

(c) Scenario 3

Figure 6.8: Comparison of the distribution of the azimuth spread of arrival for the
three scenarios in urban environment.

together with the position of the Tx and the Rx are employed to estimate the AoD
and AoA. However, during NLoS in Scenario 1, the MPCs have a short lifetime and
a weak power which results in significant phase estimation errors and consequently
cause less reliable scatterers locations estimation which may explain the errors in
the estimated path angles and angular spreads.
The results for the ASA are shown in Figure 6.8. In Scenario 1, the simulated
channels produce much larger ASA than the model input with a mean of 25.12◦
compared to 10.47◦ in the model input. The results for Scenario 2 and Scenario

128
 6.2 Large Scale Parameters

3, however, show better match. As summarized in Tables 6.1-6.3, the standard

deviation values in the model input are smaller than those of the simulated channels.
As a result, the CDFs from the model input have tails with significantly larger ASA.

Model input Simulated channels

Parameters
LoS NLoS LoS NLoS
µ 0 0 0 0
SF [dB]
σ 3.90 3.90 3.35 3.35
[Link] [m] dc 8.20 8.20 10.33 10.33
µ 15.02 N/A 14.44 N/A
K-factor [dB]
σ 4.81 N/A 4.98 N/A
[Link] [m] dc 8.01 N/A 8.83 N/A
µ -8.13 -7.44 -8.08 -7.46
DS [log10 (s)]
σ 0.34 0.16 0.31 0.19
[Link] [m] dc 8.97 49.69 9.08 50.16
µ 0.74 0.74 0.89 0.89
ASD [log10 (◦ )]
σ 0.50 0.50 0.29 0.29
[Link] [m] dc 37.11 37.11 52.04 52.04
µ 1.02 1.02 1.40 1.40
ASA [log10 (◦ )]
σ 0.40 0.40 0.25 0.25
[Link] [m] dc 48.95 48.95 55.71 55.71

Table 6.1: Comparison of LSPs for Scenario 1.

Another possible reason that could contribute to the observed deviations between
the model input and simulated channels in terms of angular spreads is the scattering
model used in the channel model implementation. As explained in Section 2.4.2,
during the procedure of the channel generation, a double-bounce model is considered.
During a channel segment, the positions of the scatterers stay fixed and are used to
update the path delays and angles at each measurement snapshot. However, due to
the limitation of the SISO measurements, a single-bounce model is considered in

129
6. Model Validation

the parameterization of the channel model where only the last-bounce scatterers are
localized.

Model input Simulated channels

Parameters
LoS OLoS LoS OLoS
µ 0 0 0 0
SF [dB]
σ 1.80 1.80 2.82 2.82
[Link] [m] dc 2.27 2.27 5.54 5.54
µ 16.22 5.34 15.90 5.93
K-factor [dB]
σ 2.88 4.38 2.87 4.80
[Link] [m] dc 4.42 30.58 4.57 46.05
µ -8.19 -7.60 -8.10 -7.62
DS [log10 (s)]
σ 0.19 0.21 0.18 0.21
[Link] [m] dc 4.42 70 4.57 59.45
µ 0.86 0.86 0.86 0.86
ASD [log10 (◦ )]
σ 0.39 0.39 0.32 0.32
[Link] [m] dc 63.23 63.23 66.03 66.03
µ 1.03 1.03 1.14 1.14
ASA [log10 (◦ )]
σ 0.39 0.39 0.26 0.26
[Link] [m] dc 66.17 66.17 71.48 71.48

Table 6.2: Comparison of LSPs for Scenario 2.

6.3 Correlation Distances

Th correlation distances of the LSPs are summarized in Tables 6.1, 6.2, and6.3.
As discussed in Section 5.2, the correlation distance is calculated from the
autocorrelation function of the evaluated LSP and used as a model input to simulate
channels that maintain the spatial correlation observed in the measured channels. It
must be noted that when the 2D correlated maps of the LSPs are generated during
channel simulation, the cross-correlations between LSP pairs are applied to the

130
 6.3 Correlation Distances

correlated maps (see Step A. in Section 2.4.2). Therefore, correlation distances of

LSPs are distorted and not anymore independent from each other. It can be noticed
that the simulated channels in all scenarios show an increase in the correlation
distances of most LSPs. This mismatch between the model input and the simulated
channels in terms of correlation distance, due to the aforementioned cross-correlation,
is inevitable. On the other hand, when a LSP shows insignificant correlation with
the other LSP, the change in its correlation distance will be small. For example, in
Scenario 3 the ASD has almost no correlation with the other LSPs in the model
input. As a result, the correlation distance calculated from the simulated channels
has an almost perfect agreement with the model input.

Model input Simulated channels

Parameters
LoS OLoS LoS OLoS
µ 0 0 0 0
SF [dB]
σ 4.69 4.69 6.31 6.31
[Link] [m] dc 7.60 7.60 16.43 16.43
µ 17.32 6.44 16.76 7.18
K-factor [dB]
σ 4.09 5.31 4.42 5.36
[Link] [m] dc 4.23 26.97 3.07 40.89
µ -8.09 -7.62 -8.07 -7.62
DS [log10 (s)]
σ 0.23 0.22 0.23 0.20
[Link] [m] dc 4.64 57.88 3.19 54.60
µ 1.04 1.04 1.02 1.02
ASD [log10 (◦ )]
σ 0.30 0.30 0.30 0.30
[Link] [m] dc 55.61 55.61 56.24 56.24
µ 1.11 1.11 1.16 1.16
ASA [log10 (◦ )]
σ 0.33 0.33 0.28 0.28
[Link] [m] dc 54.48 54.48 58.12 58.12

Table 6.3: Comparison of LSPs for Scenario 3.

131
6. Model Validation

6.4 Cross-correlation Parameters

Table 6.4 summarizes the mean cross-correlation values of LSPs pairs that were
set to the model input in comparison with the values that are calculated from the
simulated channels. These values, which correspond to the model input, are the
same as in Table 5.2. Note that, all cross-correlation matrices are positive definite.
There are 5 different LSPs which result in 10 cross-correlation values. For each
scenario, 10 values are obtained from the measurements and 10 values from the
simulated channels. The cross-correlation values measure the inter-dependency
between each pair of the LSPs and no distinction between different propagation
conditions is made for the calculation of the cross-correlation, i.e., the complete data
set of each measurement run is considered. Cross-correlation coefficients take values
from -1 to 1. A negative coefficient indicates a decrease in parameter x if parameter
y increases and vice versa. Positive coefficient means that x and y increase and
decrease simultaneously. When the coefficient is zero there is no dependency between
the two parameters.
In all scenarios, the cross-correlations between SF, KF, and DS calculated from
the simulated channels are in good agreement with the model input. The SF is
positively correlated with the KF while negatively correlated with the DS, i.e., when
the SF increases, the K-factor increases and the DS decreases. In all scenarios,
the DS and the KF have a strong negative correlation, as expected. A positive
correlation in all scenarios between the ASD and ASA is also observed in both
the model input and the simulated channels, however, a mismatch is noticed. The
ASD corss-correlations with SF, KF, and DS obtained from the measurements, i.e.,
model input, in the three scenarios are not consistent. For example, the ASD has
positive correlation of 0.28 with the SF in Scenario 1, negative correlation of -0.21
in Scenario 2, and no correlation in Scenario 3. Nevertheless, the ASD shows weak
or no dependency on the SF, KF and DS. Similarly, the ASA shows no correlation
with the KF and DS in all scenarios, while it has weak dependency on the SF in
Scenarios 2 and 3. On the other hand, the results from the simulated channels
do not show inconsistency. In all scenarios, both ASD and ASA are negatively
correlated with the SF and KF and positively correlated with the DS. However,

132
 6.4 Cross-correlation Parameters

in most cases, a significant mismatch between the model input and the simulated
channels is noticed. For example, the ASD shows no dependency on the KF in the
model input in Scenario 1 and Scenario 3 and has a positive correlation of 0.25
in Scenario 2. However, the values calculated from the simulated channels show
a strong positive correlation in all scenarios. Similar low agreement is noticed in
ASD-SF and ASD-DS correlation values. The difference in scattering model and the
estimation error of the angular spreads discussed in Section 6.2.4 can be partially
explained by the aforementioned mismatch.

Model input Simulated channels

Cross-correlation
SF KF DS ASD ASA SF KF DS ASD ASA
SF 1 0.50 -0.42 0.28 -0.07 1 0.61 -0.34 -0.07 -0.01
KF 1 -0.86 0 -0.01 1 -0.78 -0.44 -0.35
Scenario 1 DS 1 -0.33 -0.04 1 0.09 0.48
ASD 1 0.58 1 0.31
ASA 1 1
SF 1 0.35 -0.30 -0.21 -0.36 1 0.35 -0.30 -0.01 -0.11
KF 1 -0.87 0.25 -0.01 1 -0.91 -0.50 -0.63
Scenario 2 DS 1 -0.13 0.07 1 0.47 0.58
ASD 1 0.63 1 0.28
ASA 1 1
SF 1 0.63 -0.51 -0.07 -0.26 1 0.49 -0.49 -0.21 -0.26
KF 1 -0.85 0 -0.1 1 -0.82 -0.48 -0.55
Scenario 3 DS 1 0.02 -0.07 1 0.44 0.45
ASD 1 0.5 1 0.16
ASA 1 1

Table 6.4: Comparison of cross-correlation coefficients for the three scenarios in

urban environment.

133
6. Model Validation

6.5 Summary
In this chapter, The WINNER-type channel model for V2VRU communications
in critical accident scenarios is validated. The measured channels are simulated
using QuaDRiGa implementation based on the proposed model parametrization. In
addition to the path loss, the distributions of the LSPs, their correlation distances,
and the cross-correlation coefficients between all LSPs pairs are selected as validation
metrics. The LSPs are estimated from the simulated channels in the same way
they were estimated from the measured channels and then fitted to a log-normal
distribution.
It can be noted that the channel model clearly provides an almost perfect match
in terms of path loss in all three measured scenarios. Moreover, the results show
that the proposed model has a good agreement with the measurements in terms of
shadow fading in Scenario 1. However, some mismatch is observed in Scenario 2 and
Scenario 3, in which the SF extracted from the simulated channels has larger standard
deviation compared to the measured channels. The findings for the DS and the
KF show almost perfect agreements between the simulated and measured channels.
In terms of the ASD and ASA, the proposed channel model in Scenario 1 shows
some mismatch. However, better agreement can be seen in Scenario 2 and Scenario
3. Furthermore, based on the correlation distance results, it is concluded that due
to applying the cross-correlations between LSPs pairs, the correlation distances of
LSPs are not independent anymore. Therefore, LSPs, which have high correlation
values with other LSPs, experience an increase in their correlation distances. From
the cross-correlation results, it is noticed that the cross-correlations between SF,
KF, and DS calculated from the simulated channels are in good agreement with the
measured channels in all scenarios. However, the cross-correlations related to ASD
and ASA show a mismatch.
Generally, it can be concluded that the GSCM with the proposed model
parameters is able to produce fairly similar statistics as the measured channels.
The proposed model provides a good representation for the propagation channel in
V2VRU communications in the considered scenarios.

134
 Summary and Future Work
7
Vehicle-to-vulnerable road users (V2VRUs) communication overcomes the limitations
of sensor-based collision avoidance systems and provides 360◦ of awareness. In order
to develop a reliable communication system, realistic channel model in relevant
accident scenarios are of immense importance, but had yet not been thoroughly
investigated or modeled. The primary target of this thesis is to develop a geometry-
based stochastic channel model (GSCM) for V2VRUs communication.
For this purpose, two SISO channel measurement campaigns were conducted.
Both campaigns were executed using the RUSK-DLR channel sounder at a carrier
frequency of fc = 5.2 GHz and a bandwidth of B = 120 MHz. The first campaign
was executed in open-field controlled environment considering an accident scenario
between a vehicle and a pedestrian. This location made it possible to isolate and
study the impact on the received power by the different elements in the propagation
environment as well as by the mobility of the Tx and Rx. The second measurement
campaign was conducted in urban environment. The three most critical accident
scenarios involving pedestrians and cyclists were considered. The proposed channel
model in this thesis is based on data collected during this campaign.
Based on the measurement data in open-field scenarios, path loss models were
proposed. The two-ray model was found to provide a good fit to the measured

135
7. Summary and Future Work

path loss in the LoS scenario with a static tripod as a receiver. By replacing the
static tripod with a moving pedestrian, rapid fluctuations of the measured path
loss were observed. These fluctuations were found to be the result of reflections
from the body of pedestrian and the change in the antenna height due to the
pedestrian movement. The losses in power due to crowd shadowing and blockage of
the LoS by parked vehicles were also investigated and quantified. Based on the path
loss models, the shadow fading was then extracted to study its spatial correlation.
The autocorrelation functions of the measured shadow fading were calculated and
modeled. Furthermore, it is found that the most critical accident scenarios that
involve pedestrians and cyclists occur when the visibility is blocked by parked
vehicles. Therefore, motivated by this fact, a 3D ray tracing tool was developed to
detect the diffraction edges on the parked vehicles. The Fresnel-Kirchoff parameter
was then used to calculate the knife-edge diffraction loss. Moreover, based on the
measurement data in urban scenarios, the path loss was calculated and a multi-slope
log-distance path loss model was proposed for each scenario to later be used as a
channel model parameter.
Due to the non-stationarity nature of the vehicular channels, the large scale
parameters (LSPs) need to be evaluated within regions, i.e. where the wide-sense
stationary (WSS) assumption holds. Therefore, based on the measurement data in
the urban environment, the non-stationarity of the V2VRU channel was investigated.
The length of the WSS regions, i.e., stationarity distance, was obtained by estimating
the generalized local scattering function and its collinearity.
In order to estimate the LSPs, the multipath parameters were estimated using
the KEST algorithm. The time-variant CIR in the urban environment was found to
be highly cluttered by diffuse MPCs (DMCs). Therefore, a novel method to separate
specular MPCs (SMCs) and DMCs based on the density of their neighboring MPCs
was proposed. Furthermore, an algorithm for SMCs tracking based on their delay
and magnitude was presented. Having the SMCs tracked over time, each SMC is
then related to a physical scatterer in the propagation environment. Localization
of the scatterers was done using a joint delay-Doppler estimation algorithm. The
estimated positions of the scatterers were then used to estimate the AoD and AoA.

136
 A full parametrization for the WINNER-type GSCM was proposed. The LSPs
were estimated in the power and delay domain, i.e., shadow fading (SF), delay spread
(DS), narrowband K-factor (KF), and in the angular domain, i.e., azimuth angle
of departure (AoD), azimuth angle of arrival (AoA). The log-normal distribution
was found to provide a good fit to the distributions of the LSPs. Furthermore, the
spatial correlations of the LSPs were analyzed and their correlation distances were
calculated. In order to ensure the channel spatial consistency, the cross-correlation
coefficients among the LSPs were also calculated. The model parameters were then
used as an input to the QuaDRiGa channel simulator.
Finally, the proposed channel model was validated by comparing the simulated
channels with the measured channels. In addition to the path loss, the distributions
of the LSPs, their correlation distances, and the cross-correlation coefficients between
all LSPs pairs were selected as validation metrics. It can be concluded that the
GSCM with the proposed model parameters is able to produce fairly similar statistics
as the measured channels. The proposed model provides good representation for
the V2VRU propagation channel in the considered scenarios.
Future work related to this research could begin with including the DMC in the
channel model to account for their contribution to the received power. Due to time
constraints, only the measured data in one location has been analyzed and used to
model the channel. Further work is needed to analyze the measured data collected in
the other two locations in order to improve the channel model. Moreover, additional
measurement campaigns are required to cover other interesting accidents scenarios.
MIMO measurement campaigns would be valuable to improve the estimation of
angular spreads and scatterer locations.

137
 Bibliography

[1] Road traffic injuries. url: http : / / www . who . int / violence _ injury _
prevention/road_traffic/en/ (visited on 09/25/2018).
[2] M. DaSilva, J. Smith, and W. Najm. Analysis of pedestrian crashes. Tech. rep.
2003.
[3] M. Kuehn, T. Hummel, and A. Lang. “Cyclist-car accidents–their consequences
for cyclists and typical accident scenarios”. In: Proceedings of the 24th
International Conference on the Enhanced Safety of Vehicles. 2015.
[4] F. de Ponte Müller. “Survey on ranging sensors and cooperative techniques
for relative positioning of vehicles”. In: Sensors 17.2 (2017), p. 271.
[5] D. Geronimo et al. “Survey of pedestrian detection for advanced driver
assistance systems”. In: IEEE transactions on pattern analysis and machine
intelligence 32.7 (2010), pp. 1239–1258.
[6] P. Hurney et al. “Review of pedestrian detection techniques in automotive far-
infrared video”. In: IET intelligent transport systems 9.8 (2015), pp. 824–832.
[7] X. Wu et al. “Cars talk to phones: A DSRC based vehicle-pedestrian safety
system”. In: Vehicular Technology Conference (VTC Fall), 2014 IEEE 80th.
IEEE. 2014, pp. 1–7.
[8] M. Jutila et al. “ITS-G5 performance improvement and evaluation for
vulnerable road user safety services”. In: IET Intelligent Transport Systems
11.3 (2017), pp. 126–133.

139
BIBLIOGRAPHY

[9] W. Viriyasitavat et al. “Vehicular Communications: Survey and Challenges of

Channel and Propagation Models”. In: IEEE Vehicular Technology Magazine
10.2 (June 2015), pp. 55–66. issn: 1556-6072. doi: 10 . 1109 / MVT . 2015 .
2410341.
[10] D. W. Matolak. “Modeling the vehicle-to-vehicle propagation channel: A
review”. In: Radio Science 49.9 (2014), pp. 721–736.
[11] A. F. Molisch et al. “A survey on vehicle-to-vehicle propagation channels”. In:
IEEE Wireless Communications 16.6 (Dec. 2009), pp. 12–22. issn: 1536-1284.
doi: 10.1109/MWC.2009.5361174.
[12] L. Bernado et al. “Delay and Doppler spreads of nonstationary vehicular
channels for safety-relevant scenarios”. In: IEEE Transactions on Vehicular
Technology 63.1 (2014), pp. 82–93.
[13] A. Paier. “The vehicular radio channel in the 5 GHz band”. PhD thesis. 2010.
[14] R. Verdone and A. Zanella. Pervasive mobile and ambient wireless communi-
cations: COST action 2100. Springer Science & Business Media, 2012.
[15] Y. d. J. Bultitude and T. Rautiainen. “IST-4-027756 WINNER II D1. 1.2
V1. 2 WINNER II Channel Models”. In: EBITG, TUI, UOULU, CU/CRC,
NOKIA, Tech. Rep (2007).
[16] I. Rashdan et al. “Large-Scale Fading Characteristics and Models for Vehicle-to-
Pedestrian Channel at 5-GHz”. In: IEEE Access 7 (2019), pp. 107648–107658.
[17] I. Rashdan et al. “Measurement-based geometrical characterisation of the
vehicle-to-vulnerable-road-user communication channel”. In: IET Microwaves,
Antennas & Propagation (2020).
[18] I. Rashdan, F. de Ponte Muller, and S. Sand. “Analysis of Communication
Requirements for CACC in Stop-and-Go Behavior for Energy Efficient Driving”.
In: 2016 IEEE 84th Vehicular Technology Conference (VTC-Fall). 2016, pp. 1–
5. doi: 10.1109/VTCFall.2016.7880905.

140
 BIBLIOGRAPHY

[19] I. Rashdan, F. de Ponte Muller, and S. Sand. “Performance Evaluation of Traffic

Information Dissemination Protocols for Dynamic Route Planning Application
in VANETs”. In: 2016 IEEE 84th Vehicular Technology Conference (VTC-Fall).
2016, pp. 1–5. doi: 10.1109/VTCFall.2016.7881161.
[20] I. Rashdan, S. Sand, and F. de Ponte Müller. “ITS-G5 Challenges and 5G
Solutions for Vehicular Platooning”. In: WWRF37, Kassel, Germany. Jan.
2016.
[21] I. Rashdan et al. “Performance Evaluation of Vehicle-to-Vehicle Communica-
tion for Cooperative Collision Avoidance at Urban Intersections”. In: 2017
IEEE 86th Vehicular Technology Conference (VTC-Fall). Sept. 2017, pp. 1–5.
doi: 10.1109/VTCFall.2017.8288274.
[22] I. Rashdan et al. “Vehicle-to-Pedestrian Channel Characterization: Wide-
band Measurement Campaign and First Results”. In: IEEE 12th European
Conference on Antennas and Propagation (EuCAP). London, UK, Apr. 2018.
[23] I. Rashdan et al. “Large Scale Fading Characteristics for Vehicle-to-Cyclist
Channel in Urban Environment at 5 GHz”. In: 2020 14th European Conference
on Antennas and Propagation (EuCAP). 2020, pp. 1–5. doi: 10 . 23919 /
EuCAP48036.2020.9135985.
[24] I. Rashdan et al. “Measurement-based Geometrical Characterization of
the Vehicle-to-Pedestrian Channel”. In: 2019 13th European Conference on
Antennas and Propagation (EuCAP). IEEE. 2019, pp. 1–5.
[25] I. Rashdan et al. “Diffuse Multipath Analysis of Vehicle-to-Vulnerable Road
User Channel in Urban Environment”. In: 2021 15th European Conference
on Antennas and Propagation (EuCAP). 2021, pp. 1–5. doi: 10 . 23919 /
EuCAP51087.2021.9411391.
[26] M. F. Iskander and Z. Yun. “Propagation prediction models for wireless
communication systems”. In: IEEE Transactions on microwave theory and
techniques 50.3 (2002), pp. 662–673.
[27] J. D. Parsons and P. J. D. Parsons. The mobile radio propagation channel.
Vol. 2. Wiley New York, 2000.

141
BIBLIOGRAPHY

[28] O. Landron, M. J. Feuerstein, and T. S. Rappaport. “A comparison of

theoretical and empirical reflection coefficients for typical exterior wall surfaces
in a mobile radio environment”. In: IEEE Transactions on Antennas and
Propagation 44.3 (1996), pp. 341–351.
[29] P. Beckmann and A. Spizzichino. “The scattering of electromagnetic waves
from rough surfaces”. In: ah (1987).
[30] S. Salous. Radio propagation measurement and channel modelling. John Wiley
& Sons, 2013.
[31] P. Series. “Propagation by diffraction”. In: Recommendation ITU-R P. 526-13
(2013).
[32] S. Priebe, M. Jacob, and T. Kürner. “Calibrated broadband ray tracing for the
simulation of wave propagation in mm and sub-mm wave indoor communication
channels”. In: European Wireless 2012; 18th European Wireless Conference
2012. VDE. 2012, pp. 1–10.
[33] H. C. Nguyen et al. “Evaluation of empirical ray-tracing model for an urban
outdoor scenario at 73 GHz E-band”. In: 2014 IEEE 80th Vehicular Technology
Conference (VTC2014-Fall). IEEE. 2014, pp. 1–6.
[34] T. Huschka. “Ray tracing models for indoor environments and their compu-
tational complexity”. In: 5th IEEE International Symposium on Personal,
Indoor and Mobile Radio Communications, Wireless Networks-Catching the
Mobile Future. Vol. 2. IEEE. 1994, pp. 486–490.
[35] T. Fugen et al. “Capability of 3-D ray tracing for defining parameter sets
for the specification of future mobile communications systems”. In: IEEE
Transactions on antennas and propagation 54.11 (2006), pp. 3125–3137.
[36] G. Acosta-Marum. “Measurement, modeling, and OFDM synchronization for
the wideband mobile-to-mobile channel”. PhD thesis. Georgia Institute of
Technology, 2007.
[37] I. S. Association et al. “Wireless lan medium access control (mac) and
physical layer (phy) specifications amendment 6: Wireless access in vehicular
environments”. In: Standard IEEE 802.11 (2010).

142
 BIBLIOGRAPHY

[38] P. Heino et al. WINNER+ final channel models. Tech. rep. Tech. Rep, 2010.
[39] 3. T. 3. v12.2.0. Study on 3D channel model for LTE. Tech. rep. Tech. Rep,
2015.
[40] S. Jaeckel. “Quasi-deterministic channel modeling and experimental validation
in cooperative and massive MIMO deployment topologies”. PhD thesis.
Dissertation, Ilmenau, TU Ilmenau, 2017, 2017.
[41] C. Gustafson et al. “The COST IRACON Geometry-Based Stochastic Channel
Model for Vehicle-to-Vehicle Communication in Intersections”. In: IEEE
Transactions on Vehicular Technology 69.3 (2020), pp. 2365–2375. doi: 10.
1109/TVT.2020.2964277.
[42] H. Yi et al. “Characterization for the Vehicle-to-Infrastructure Channel in
Urban and Highway Scenarios at the Terahertz Band”. In: IEEE Access 7
(2019), pp. 166984–166996. doi: 10.1109/ACCESS.2019.2953890.
[43] G. Makhoul, R. D’Errico, and C. Oestges. “Wideband Vehicle to Pedestrian
Propagation Channel Characterization and Modeling”. In: IEEE 12th European
Conference on Antennas and Propagation (EuCAP). London, UK, Apr. 2018.
[44] G. Makhoul, R. D’Errico, and C. Oestges. “Wideband Measurement-Based
Vehicle-to-Pedestrian Channel Models”. In: IEEE Transactions on Vehicular
Technology 68.10 (2019), pp. 9339–9347. doi: 10.1109/TVT.2019.2936629.
[45] G. Makhoul, R. D’Errico, and C. Oestges. “Characterization of Vehicle to
Pedestrian Correlation Properties”. In: 2020 IEEE International Symposium
on Antennas and Propagation and North American Radio Science Meeting.
2020, pp. 1113–1114. doi: 10.1109/IEEECONF35879.2020.9330040.
[46] M. G. Doone and S. L. Cotton. “First-order characteristics of the person-to-
vehicle channel at 5.8 GHz”. In: Antennas and Propagation (EuCAP), 2014
8th European Conference on. IEEE. 2014, pp. 848–851.
[47] M. G. Doone et al. “Pedestrian-to-vehicle communications in an urban
environment: Channel measurements and modeling”. In: IEEE Transactions
on Antennas and Propagation 67.3 (2018), pp. 1790–1803.

143
BIBLIOGRAPHY

[48] J. J. Anaya et al. “Vehicle to pedestrian communications for protection

of vulnerable road users”. In: 2014 IEEE Intelligent Vehicles Symposium
Proceedings. June 2014, pp. 1037–1042. doi: 10.1109/IVS.2014.6856553.
[49] Z. Liu et al. “Implementation and performance measurement of a V2X
communication system for vehicle and pedestrian safety”. In: International
Journal of Distributed Sensor Networks 12.9 (2016), p. 1550147716671267.
[50] P. Sewalkar, S. Krug, and J. Seitz. “Towards 802.11 p-based vehicle-to-
pedestrian communication for crash prevention systems”. In: Ultra Modern
Telecommunications and Control Systems and Workshops (ICUMT), 2017 9th
International Congress on. IEEE. 2017, pp. 404–409.
[51] M. Bagheri, M. Siekkinen, and J. K. Nurminen. “Cloud-based pedestrian
road-safety with situation-adaptive energy-efficient communication”. In: IEEE
Intelligent transportation systems magazine 8.3 (2016), pp. 45–62.
[52] M. Bagheri, M. Siekkinen, and J. K. Nurminen. “Cellular-based vehicle
to pedestrian (V2P) adaptive communication for collision avoidance”. In:
Connected Vehicles and Expo (ICCVE), 2014 International Conference on.
IEEE. 2014, pp. 450–456.
[53] D. S. Baum et al. “An interim channel model for beyond-3G systems: extending
the 3GPP spatial channel model (SCM)”. In: 2005 IEEE 61st Vehicular
Technology Conference. Vol. 5. IEEE. 2005, pp. 3132–3136.
[54] K. Bakowski and K. Wesolowski. “Change the channel”. In: IEEE Vehicular
Technology Magazine 6.2 (2011), pp. 82–91.
[55] G. L. Turin et al. “A statistical model of urban multipath propagation”. In:
IEEE Transactions on Vehicular Technology 21.1 (1972), pp. 1–9.
[56] T. S. Rappaport et al. Wireless communications: principles and practice. Vol. 2.
prentice hall PTR New Jersey, 1996.
[57] H. Fernández et al. “Path loss characterization for vehicular communications at
700 MHz and 5.9 GHz under LOS and NLOS conditions”. In: IEEE Antennas
and Wireless Propagation Letters 13 (2014), pp. 931–934.

144
 BIBLIOGRAPHY

[58] M. Luo, G. Villemaud, and J.-M. Gorce. “Generation of 2D correlated random

shadowing based on the deterministic MR-FDPF model”. In: EURASIP
Journal on Wireless Communications and Networking 2015.1 (2015), p. 208.
[59] I. Forkel, M. Schinnenburg, and M. Ang. “Generation of two-dimensional
correlated shadowing for mobile radio network simulation”. In: WPMC, sep
21 (2004), p. 43.
[60] T. Abbas et al. “A measurement based shadow fading model for vehicle-
to-vehicle network simulations”. In: International Journal of Antennas and
Propagation 2015 (2015).
[61] T. W. Beck et al. “An examination of the runs test, reverse arrangements
test, and modified reverse arrangements test for assessing surface EMG signal
stationarity”. In: Journal of neuroscience methods 156.1-2 (2006), pp. 242–248.
[62] M. Gudmundson. “Correlation model for shadow fading in mobile radio
systems”. In: Electronics letters 27.23 (1991), pp. 2145–2146.
[63] T. B. Sorensen. “Correlation model for slow fading in a small urban macro
cell”. In: Personal, Indoor and Mobile Radio Communications, 1998. The
Ninth IEEE International Symposium on. Vol. 3. IEEE. 1998, pp. 1161–1165.
[64] Y. Zhang et al. “A novel spatial autocorrelation model of shadow fading in
urban macro environments”. In: Global Telecommunications Conference, 2008.
IEEE GLOBECOM 2008. IEEE. IEEE. 2008, pp. 1–5.
[65] K. D. Mielenz. “Computation of Fresnel integrals. II”. In: Journal of research
of the National Institute of Standards and Technology 105.4 (2000), p. 589.
[66] J. Epstein and D. W. Peterson. “An experimental study of wave propagation
at 850 Mc”. In: Proceedings of the IRE 41.5 (1953), pp. 595–611.
[67] A. Paier et al. “Non-WSSUS vehicular channel characterization in highway
and urban scenarios at 5.2 GHz using the local scattering function”. In: 2008
international itg workshop on smart antennas. IEEE. 2008, pp. 9–15.
[68] G. Matz. “On non-WSSUS wireless fading channels”. In: IEEE transactions
on wireless communications 4.5 (2005), pp. 2465–2478.

145
BIBLIOGRAPHY

[69] A. Ispas et al. “Analysis of local quasi-stationarity regions in an urban macrocell

scenario”. In: 2010 IEEE 71st Vehicular Technology Conference. IEEE. 2010,
pp. 1–5.
[70] D. Slepian. “Prolate spheroidal wave functions, Fourier analysis, and uncer-
tainty—V: The discrete case”. In: Bell System Technical Journal 57.5 (1978),
pp. 1371–1430.
[71] R. Thomä, M. Landmann, and A. Richter. “RIMAX—a maximum likelihood
framework for parameter estimation in multidimensional channel sounding”.
In: Proceedings of the International Symposium on Antennas and Propagation
(ISAP’04). Citeseer. 2004, pp. 53–56.
[72] B. H. Fleury et al. “Channel parameter estimation in mobile radio environ-
ments using the SAGE algorithm”. In: IEEE Journal on selected areas in
communications 17.3 (1999), pp. 434–450.
[73] T. Jost et al. “Detection and tracking of mobile propagation channel paths”. In:
IEEE Transactions on Antennas and Propagation 60.10 (2012), pp. 4875–4883.
[74] T. Jost et al. “Using an autoregressive model for DMC”. In: 2012 6th European
Conference on Antennas and Propagation (EUCAP). IEEE. 2012, pp. 3504–
3508.
[75] N. Savage et al. “Radio wave propagation through vegetation: Factors
influencing signal attenuation”. In: Radio Science 38.5 (2003), pp. 9–1.
[76] M. Al-Nuaimi and A. Hammoudeh. “Measurements and predictions of
attenuation and scatter of microwave signals by trees”. In: IEE Proceedings-
Microwaves, Antennas and Propagation 141.2 (1994), pp. 70–76.
[77] N. Schneckenburger et al. “Reflector Localization for Geometrical Modeling
the Air-Ground Channel”. In: IEEE Transactions on Vehicular Technology
(2018).
[78] A. Richter. “Estimation of radio channel parameters: Models and algorithms”.
In: ISLE. 2005.

146
 BIBLIOGRAPHY

[79] J. Poutanen et al. “Angular and shadowing characteristics of dense multipath

components in indoor radio channels”. In: IEEE Transactions on Antennas
and Propagation 59.1 (2010), pp. 245–253.
[80] A. Richter, J. Salmi, and V. Koivunen. “Distributed scattering in radio channels
and its contribution to MIMO channel capacity”. In: 2006 First European
Conference on Antennas and Propagation. IEEE. 2006, pp. 1–7.
[81] G. Calcev et al. “A wideband spatial channel model for system-wide
simulations”. In: IEEE Transactions on Vehicular Technology 56.2 (2007),
pp. 389–403.
[82] M. Series. “Guidelines for evaluation of radio interface technologies for IMT-
Advanced”. In: Report ITU 638 (2009), pp. 1–72.
[83] P. Kyosti. “WINNER II channel models”. In: IST, Tech. Rep. IST-4-027756
WINNER II D1. 1.2 V1. 2 (2007).
[84] E. Tanghe et al. “Experimental analysis of dense multipath components in an
industrial environment”. In: IEEE Transactions on Antennas and Propagation
62.7 (2014), pp. 3797–3805.
[85] B. Hanssens et al. “An extension of the RiMAX multipath estimation algorithm
for ultra-wideband channel modeling”. In: EURASIP journal on wireless
communications and networking 2018.1 (2018), p. 164.
[86] S. R. Saunders and A. Aragón-Zavala. Antennas and propagation for wireless
communication systems. John Wiley & Sons, 2007.
[87] L. J. Greenstein, D. G. Michelson, and V. Erceg. “Moment-method estimation
of the Ricean K-factor”. In: IEEE communications letters 3.6 (1999), pp. 175–
176.
[88] v. 3GPP TR 25.996. Spatial channel model for multiple input multiple output
(MIMO) simulations. Tech. rep. 3rd Generation Partnership Project (3GPP),
2003.

147