See discussions, stats, and author profiles for this publication at: [Link]

net/publication/319271953

Applications of Trajectory Data From the Perspective of a Road Transportation

Agency: Literature Review and Maryland Case Study

Article  in  IEEE Transactions on Intelligent Transportation Systems · June 2018

DOI: 10.1109/TITS.2018.2843298

CITATIONS READS

22 506

5 authors, including:

Nikola Markovic Przemysław Sekuła

University of Maryland, College Park University of Maryland, College Park
22 PUBLICATIONS   219 CITATIONS    11 PUBLICATIONS   63 CITATIONS   

SEE PROFILE SEE PROFILE

Gennady Andrienko
Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS
335 PUBLICATIONS   12,243 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Variety, Veracity, VaLue: Handling the Multiplicity of Urban Sensors View project

Track and Know, EU-Horizon 2020 project on "Big Data for Mobility Tracking Knowledge Extraction in Urban Areas" View project

All content following this page was uploaded by Nikola Markovic on 01 June 2018.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 1

Applications of Trajectory Data from the Perspective

of a Road Transportation Agency: Literature Review
and Maryland Case Study
Nikola Marković, Przemysław Sekuła, Zachary Vander Laan, Gennady Andrienko, and Natalia Andrienko

Abstract—Transportation agencies have an opportunity to of a transportation agency, seeking to synthesize existing

leverage increasingly-available trajectory datasets to improve approaches and also present new applications for transportation
their analyses and decision-making processes. However, this data systems analysis. It is worth noting that a recent review paper
is typically purchased from vendors, which means agencies must [1] also seeks to bring trajectory data closer to practice;
understand its potential benefits beforehand in order to properly however, it focuses in particular on visual analytics approaches
assess its value relative to the cost of acquisition. While the
literature concerned with trajectory data is rich, it is naturally
that may be useful for transportation agencies. The authors
fragmented and focused on technical contributions in niche areas, conclude that it is necessary to establish collaboration between
which makes it difficult for government agencies to assess its value the visual analytics and transportation research communities,
across different transportation domains. To overcome this issue, and seek to do so in the current paper.
the current paper explores trajectory data from the perspective of The trajectories analyzed in this paper were obtained from
a road transportation agency interested in acquiring trajectories a major GPS company in North America. It provides Internet
to enhance its analyses. The paper provides a literature review services and mobile applications informing users about traffic
illustrating applications of trajectory data in six areas of road conditions, which are estimated based on terabytes of GPS
transportation systems analysis: demand estimation, modeling data collected daily from millions of mobile phones, cars,
human behavior, designing public transit, traffic performance
trucks and other fleet vehicles. In 2016 the Maryland State
measurement and prediction, environment and safety. In addition,
it visually explores 20 million GPS traces in Maryland, illustrating Highway Administration (SHA) purchased GPS traces of all
existing and suggesting new applications of trajectory data. trips recorded in Maryland during four months of the previous
year. The SHA subsequently asked the authors of this paper
Keywords—road transportation, trajectory data, literature review, to determine the value of the trajectory data for transportation
visual analytics, machine learning, big data. systems analysis and evaluate the cost/benefit trade-off. With
the goal of enabling SHA and other government agencies
I. I NTRODUCTION to accurately assess the value of these datasets, this paper
provides an overview of potential use-cases in various domains
Numerous detailed trajectory datasets have recently become of transportation engineering. In particular, we make two
available, including Global Positioning System (GPS) traces contributions:
from cell phones and vehicles, anonymized Call Detail Records
(CDR) from cell phone providers, and data from arrays of • We provide a literature review illustrating innovative
Bluetooth and Wi-Fi detectors that re-identify devices over uses of trajectory data in road transportation systems
time. As vast amounts of spatiotemporal data becomes more analysis. The review includes studies that exploit dif-
ubiquitous, transportation agencies have an opportunity to ferent trajectory datasets (GPS traces, CDR, Bluetooth
leverage these resources to improve analysis techniques and and Wi-Fi detectors) in six areas of transportation engi-
answer important questions more efficiently. However, since neering: demand estimation, modeling human behavior,
this data often needs to be purchased, agencies should be designing public transit, traffic performance measure-
well-informed about potential benefits in order to assess its ment and prediction, environmental impact, and safety
value to their organization. While the literature concerned analysis. This review can serve as a single reference
with trajectory data is rich, it is naturally fragmented and point for government agencies trying to decide whether
often focuses on technical contributions in niche areas, which purchasing trajectory data would be beneficial to their
makes it hard for government agencies to assess its specific multifaceted analyses.
application to transportation domains. To overcome this issue, • We visually explore a set of 20 million GPS traces in
the current paper explores trajectory data from the perspective Maryland, demonstrating existing and suggesting new
applications of trajectory data in road transportation
The first three authors are with the Center for Advanced Transportation systems analysis. The suggested novel applications in-
Technology, Department of Civil and Environmental Engineering, University clude: (a) design of isochrones via density-based clus-
of Maryland, College Park, MD, USA, while the last two authors are with the tering/filtering of trajectory data that can be applied
Fraunhofer Institute for Intelligent Analysis and Information Systems, Sankt
Augustin, Germany and the City University London, UK. The second author without any information about the underlying transporta-
is also affiliated with the University of Economics in Katowice, Poland. tion network and historical travel times along different
Manuscript received December 7, 2017, revised April 6, 2018; road links, and (b) weight/speed enforcement that could
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 2

improve safety and reduce property damage while em- of time. The region is then divided into a set of zones, and
ploying simple processing and visualization of trajectory stay points are assigned to the zone that encompasses them,
data. Lastly, we summarize the best-practices in ana- meaning that the stay-point trajectories represent trips between
lyzing trajectories and discuss data-related challenges zones. After carefully discarding users who do not use their
that transportation agencies should be aware of when phones frequently enough to accurately characterize their travel
purchasing data. behavior and scaling the results to reflect the total population,
The next section provides a literature review illustrating an estimate of daily O-D trip tables can be produced.
various uses of trajectory data in road transportation, while 2) Additional considerations: It should be noted that other
the following section showcases its application in Maryland. demand models consider time-of-day effects (e.g., AM/PM
After discussing data-related challenges, we conclude by sum- Peak, Off Peak) and the types of trips taken (e.g., Home-Work,
marizing the findings. Home-Other), which can also be extracted from mobility data.
These considerations will be discussed in the next subsection.
II. L ITERATURE R EVIEW
Applications of trajectory data in road transportation are B. Modeling human behavior
synthesized into six areas, each of which is discussed in Quantifying human behavior is a key component of demand
a separate subsection. It is worth noting that the following modeling and transportation planning, since understanding
review does not seek to provide an exhaustive overview of why people travel and the specific choices they make in the
the literature, but highlight some of the relevant work in order process (e.g., mode and route choice) can be useful for shaping
to illustrate applications of trajectory data in different areas policies that positively impact the overall transportation sys-
of transportation engineering. The review generally focuses tem. This subsection focuses on two specific aspects of human
on relatively recent papers that include comprehensive case behavior: assigning context to travel movements and choice
studies, which could be of particular interest to transportation analysis.
agencies. 1) Context of travel movement: Given detailed mobility
datasets, intelligent data mining strategies can be utilized to
derive meaning and context from the locations visited. A
A. Demand estimation recent paper [5] provides a thorough overview of the field,
At the core of demand modeling and transportation planning distilling trajectory data mining into the following phases:
is the problem of estimating the number of trips that take (a) preprocessing (trajectory compression, stay-point detection,
place between specific locations [2]. The traditional data trajectory segmentation and map matching), (b) data manage-
sources used to estimate demand are census and travel sur- ment (indexing and storing data so it can be retrieved quickly)
vey data, which are sometimes combined with traffic counts and (c) pattern mining (clustering by time/shape/segment, clas-
from roadside sensors. While these datasets contain valu- sifying, and detecting outliers). The last phase is particularly
able information, their use is sometimes hindered by non- interesting for transportation, because its application involves
representative sampling and misreported responses on surveys, grouping similar trip origins, destinations, times of day, trip
as well as difficulties with reconstructing the trips between durations, and sections of road, in order to extract prevailing
Origin-Destination (O-D) pairs based on sparse vehicle count patterns and answer transportation-related questions.
data. Given the importance of determining O-D pairs and the The aspect of trajectory mining most relevant to demand
shortcomings of traditional methods, mobility data offers an estimation is the stay-point detection process, which helps
appealing, more direct approach to inferring demand. identify locations at which an individual spends significant
1) An example O-D matrix derivation based on trajectories: amounts of time. For example, [6] uses trajectory data to
The methodology employed in [3] is an innovative example detect and classify significant locations (e.g., home, work,
of how trajectory data can be utilized to estimate demand. or social) in a way that respects user privacy, employing a
The proposed approach begins with a preprocessing procedure, visual analytics approach and demonstrating its capabilities on
where CDR data points representing timestamps of phone calls a benchmark dataset and location data from Twitter. Another
and text messages are mapped to locations, either through example includes [7], which utilizes the concept of network
triangulation methods or simply by locating the nearest cell motifs to investigate and describe the types of locations where
tower. Individual anonymous users’ locations are then tracked cell phone users spend extended periods of time.
over time to form trajectories through space, making the data 2) Choice analysis: The other aspect of modeling human
functionally similar to GPS trajectory data, but with less spatial behavior that can be improved through mobility data is de-
resolution. From this point, the goal is to mine the data to scribing choice behavior (e.g., mode and route choice). In
extract the number of trips that take place between locations, transportation, people’s behavior is usually addressed with
a process that involves making assumptions about how to discrete choice models, where users consider a set of mutually
define important locations and assign meaning to the set of exclusive alternatives and choose the one that maximizes their
movements over time. The authors use an algorithm from [4] utility. Given a set of observations about travel behavior from
to transform the detailed trajectory data into more manageable some segment of the population, a transportation modeler
trajectories of stay locations, where a stay location represents seeks to find model parameters that best describe the observed
a place in which a cell-phone user spends significant amounts behavior [8]. Consequently, detailed travel survey data is vital
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 3

to discrete choice modeling, but is laborious to acquire and optimal transit routes. One of the most complete examples of
may become outdated after a few years. Accordingly, mobility this approach is found in [11], where the authors propose a
data provides an opportunity to observe how people behave, methodology to design a new transit network in Abidjan, Ivory
from which discrete choice models can be calibrated, verified, Coast, using the aforementioned cell phone data. The premise
or shown to be flawed. This can be illustrated through the is based on the idea that a transit network’s service should
following two studies. reflect the spatial and temporal patterns of people’s movement.
A recent work [9] combines CDR, Waze GPS data and Based on patterns that emerge from the massive amount of cell
a handful of other sources to investigate the impact of spe- phone data points (referred to as m-trails), a set of potential
cial events on a city’s travel patterns, focusing on the 2016 routes are selected and then refined by employing other utility-
Olympics in Rio de Janeiro, Brazil. The authors estimate maximization strategies. Upon selecting the routes, the authors
O-D demand prior to the Olympics, and creatively utilize use linear programming to find optimal service frequencies.
the Olympic event schedule, stadium capacities, Airbnb and
hotel information to account for additional destinations and
demand from tourists. After building the demand model to D. Traffic performance measurement and prediction
account for the Olympics, they explore choice behavior in Trajectory data can be used both to analyze historical
the form of mode shift and traffic routing strategies, noting performance of a traffic system and to help predict future traffic
the overall system implications associated with the different states. Upon discussing the related work, we point out existing
choice behaviors. Another example is [10], where the authors challenges in this area.
use GPS traces of 526 vehicles to investigate routing behavior 1) Quantifying past performance: Transportation agencies
and check whether people take the lowest-cost paths, which is require traffic data in order to quantify system performance, in-
commonly assumed in traffic assignment. They cluster origins form policy decisions, and identify areas of improvement [14].
and destinations to find important locations, cluster trajectories Important performance indicators include congestion-related
to determine possible routes, and discover that most users take measures, such as travel times over different time periods,
the same path in the majority of situations, which often is travel time reliability, vehicle/person throughput, occupancy,
not the minimum cost path. Studies like these help determine and total vehicle delay, all of which depend on the ability
whether choice models that are based on utility maximization to accurately capture data. Traditional traffic sensors such as
actually match real-world behavior. induction loop detectors and radar/microwave detectors are
useful for obtaining vehicle counts, but have more difficulty
C. Designing public transit estimating travel time distributions because these fixed sensors
Public transit systems provide an effective way to help measure only spot-mean speed [15]. There are many intelligent
relieve congestion, reduce emissions, and transport people techniques that can be used to overcome this drawback of
efficiently in areas where significant travel demand exists traditional detector data, but trajectory datasets offer an al-
between common origins or destinations [11]. Transit plan- ternative, direct approach for measuring travel times. Rather
ning consists of selecting system characteristics (e.g., station than inferring travel times based on point measurements and
locations, routes, fleet size, service frequencies, fares) in order constant-speed assumptions, these datasets can be used directly
to provide satisfactory service at minimal cost [12]. This task to calculate travel time distributions, quantify congestion mea-
can be aided by trajectory data in different ways. sures, and serve as a ground truth for other sensor data [16].
1) Trajectories as input to optimization models: The tra- State and local agencies can leverage these probe vehicle data
ditional transportation network optimization techniques rely and existing methodologies to develop mobility reports.
on aggregate O-D matrices [11], which we have already 2) Real-time predictions: In addition to quantifying past
discussed in the demand estimation subsection of this paper. performance of a transportation network, traffic data can be
We reemphasize that, in addition to traditional survey/land used for real-time traffic state predictions, provided that data
use/traffic count methods, these O-D matrices can be estimated feeds are available in real time. With some exceptions (e.g.,
by mining trajectory data. Note that this approach may be [17]), literature in this area tends to focus on data assimilation
particularly useful in developing countries where survey data techniques, which seek to optimally blend predictions from
may not be available, and in cities where travel survey data traffic models and field measurement observations, each of
quickly becomes outdated due to rapid population growth. which contain some unknown levels of uncertainty [18]. While
In such cases, trajectory data may help provide reasonable significant data assimilation research has been performed using
aggregate demand estimates to feed existing transit network stationary sensors, trajectory-based measurements provide new
optimization models. An example of this approach is found in opportunities for traffic state estimation. For example, [19]
[13], where the authors use CDR to propose route changes to investigates the performance of a Kalman filtering approach
a transit system in Abidjan, Ivory Coast, resulting in estimated to travel time estimation using data from a fleet of GPS-
average travel time reductions of up to 10% across the city. enabled probe vehicles, with traditional traffic sensors serving
2) Data-driven approach: There is another, more data- as ground truth measurements. Recognizing that GPS and loop
driven approach to transit planning. Rather than reducing detector data sources contain complementary information, oth-
trajectory data to a set of important O-D locations and using ers consider assimilation techniques that merge data collected
these O-D matrices to feed an optimization model, the data- from both fixed and moving measurements, while focusing on
driven approach seeks to use the trajectory data directly to infer different aspects of the assimilation problem and application
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 4

areas (e.g., [20] considers arterial traffic in the context of network conditions after a 2011 earthquake in Japan and
disruptive events). makes recommendations for disaster management. In response
to the devastating earthquake, [28] develops a methodology
for probabilistically modeling human movement using GPS
E. Environment traces to help better respond to future disasters. Similarly, [29]
The transportation sector was responsible for 26% of all determines optimal evacuation routes after natural disasters,
2014 greenhouse gas emissions in the United States, mostly employing a multi-objective genetic algorithm to jointly opti-
from burning fossil fuel for vehicles, trains, planes, and ships mize evacuation distance, time, and safety.
[21]. Thus, transportation agencies are often interested in (a) 2) Cyclist and pedestrian safety: A separate branch of
quantifying their environmental impact, and (b) developing safety research leverages GPS, Wi-Fi and Bluetooth trajectory
strategies to make operations more efficient and shift reliance data to provide insight into cyclist and pedestrian safety.
away from fossil fuels. Both can be aided by the use of For example, [30] investigates bicycle risk by analyzing GPS
trajectory data. traces, calculating incident rates through simple odds ratios,
1) Quantifying emissions: An important way to quantify the and concluding that crash risk is greatest at intersections
environmental impact of traffic is through transportation emis- and on roads that are in poor condition. A related research
sions models, which can be approached from macroscopic or combines GPS traces with bicycle count data to infer high-
microscopic vantage points. Macro-level models (e.g., EMEP, risk areas for cycling injuries [31]. These analyses provide
EEA) base the emissions calculations on aggregate flows and methodological frameworks and recommendations that may be
average vehicle speeds along transportation networks [22]. useful for transportation agencies looking to design bike lanes
In contrast, micro-level models focus on individual vehicles’ or improve bikeshare safety.
accelerations and decelerations, which produce more accurate From a pedestrian and urban planning perspective, [32]
emissions estimates than macroscopic models, an example of uses GPS traces to characterize human movement in order
which includes VT-Micro [23]. Since trajectory data can be to address the issue of excessive pedestrian density during
used to improve demand estimation techniques, its applica- special religious events in Saudi Arabia. Likewise, [33] an-
tion to macroscopic emissions modeling yields more accurate alyzes trajectories during a crowd disaster to characterize
estimates. Similarly, since micro-level emissions models rely how pedestrian dynamics change from low to unsafe crowd
on knowledge of vehicle accelerations, trajectory data can be densities. Although the empirical data is extracted from video,
used to calculate these inputs directly rather than relying on it is nonetheless trajectory data that can be treated similarly to
estimates from microsimulation experiments (e.g., [24]). From datasets collected from other technologies.
either perspective, trajectory data provides an opportunity to
better quantify existing emissions resulting from transportation III. M ARYLAND C ASE S TUDY
operations. In this section we showcase several applications of trajectory
2) Mitigating emissions: One attempt to reduce green house data in road transportation. Most importantly, we propose three
gas emissions is by developing vehicles which use alternative innovative applications that (to the best of our knowledge) have
energy sources. Electric vehicles are one such alternative, not been considered in the literature: measuring accessibility
but are hindered by a lack of necessary infrastructure for via density-based clustering/filtering of waypoints, identifying
conveniently recharging. Thus, in an attempt to promote adop- candidate locations for speed cameras, and selecting regions
tion of electric vehicle and related technologies, cities and for additional vehicle weight enforcement. In addition, we
planning agencies may be interested in determining how to illustrate applications of trajectory data in estimating demand
best locate recharging/refueling infrastructure. A handful of and evaluating transit systems that have been extensively
recent studies suggest that trajectory data may be beneficial addressed in the literature (see Sections II-A and II-C). These
for achieving these goals, including [25]. This work uses are included because they are highly applicable to many trans-
taxi GPS traces from China as input to a facility location portation agencies, and help illustrate how the results can be
model, seeking to determine optimal locations and capacities effectively communicated to practitioners. Also, we note that
of charging facilities. Another attempt to reduce emissions is to analysis of demand is relevant for other applications discussed
use trajectory data to enable efficient carpooling [26], which in the literature review, which highlights the importance of
can be done by extracting mobility patterns from data and inferring overall traffic volumes from raw trajectory data.
using those in an optimization setting to minimize the number
of cars needed for carpooling.
A. Data
The dataset used in this paper consists of GPS trajectories
F. Safety from 20 million trips recorded during February, June, July
Trajectory data has recently been used in a number of and October of 2015. Each trip consists of an origin and
innovative applications focusing on emergency response and destination, as well as a number of intermediate waypoints,
cyclist/pedestrian safety. each of which has a corresponding time stamp (see Figure
1) Emergency response: A recent paper [27] demonstrates 1 for a sample trip). Insight into the dataset is provided
how trajectory data may be useful during emergencies by by summarizing characteristics of trips recorded during the
using probe vehicle and smartphone GPS data to assess month of October. Namely, the median trip duration and length
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 5

Trip Duration Trip Length

1) Preprocessing: Since GPS data includes measurement

Percentile of Data Set

Percentile of Data Set

13
14 80 80

60 60
errors, the recorded waypoints are not necessarily located along
40 median: 17.7 min 40 median: 7.0 mi
the physical road network. In addition, the granularity of data
20
average: 35.5 min
20
average: 23.4 mi is not always high enough to include a waypoint along every
st. deviation: 52.1 min st. deviation: 46.8 mi
2
0
0 20 40 60 80 100
0
0 20 40 60
single road link (or a traffic message channel) that a vehicle tra-
1
0 Trip Duration [min] Trip Length [mi] verses. Therefore some preprocessing is needed in order to map
match waypoints to the road network and reconstruct road-
Fig. 1: A sample trip with relatively few waypoints and based routes. This was done using the OpenStreetMap [34]
descriptive statistics for 6.4 million trips recorded in October. routing tool (Figure 4), which applies a hidden Markov model
Trip lengths are computed based on great-circle distances to find the most likely road-based route from a time-stamped
between waypoints. sequence of latitude/longitude pairs [35]. The computationally-
intensive map matching was carried out in parallel on a 10-
Time Between Waypoints Spacing Between Waypoints core computer, and took about 3 days to process all 20 million
Percentile of Data Set

Percentile of Data Set

80 80 trips. Since map matching results in trajectories that include a
60 60 significant amount of additional information (i.e., data about
40 median: 1.0 sec 40 median: 28.2 m every road link that a vehicle traverses), the corresponding
average: 17.0 sec average: 291.3 m
20
st. deviation: 34.4 sec
20
st. deviation: 587.4 m
dataset increased in size from the initial 112 GB to over
0
0 20 40 60 80
0
0 500 1000 1500
5 TB. However, after removing redundant information (i.e.,
Time [sec] Distance [m] keeping only one node per road link), the remaining dataset
Fig. 2: October waypoints and statistics computed based on a was reduced from 5 TB to 700 GB.
sample of over 360 million waypoints after removing outliers 2) Database: In order to efficiently store and query the large
(e.g., unrealistic displacements due to device-related errors). dataset we utilized PostgreSQL 9.6, an open-source database
Spacing is expressed in great-circle distances. that has several useful features for analyzing spatio-temporal
data. First, it comes with PostGIS spatial database extender,
which adds support for geographic objects and allows location
queries to be run in the Structured Query Language (SQL). It
are about 18 min and 7 miles (Figure 1), while the median is also highly integrated with QGIS, which is an open source
time lapse and spacing between consecutive waypoints are Geographic Information System (GIS) that was extensively
approximately 1 second and 28 meters respectively (Figure used in this study. Additionally, it includes a number of built-
2). About 77% of the trips are internal to Maryland, while the in solutions to facilitate data manipulations, such as table
remaining 23% have at least one waypoint outside Maryland inheritance mechanism, spatial indexing, and advanced spatial
(Figure 3). The same visual indicates that the vast majority of queries. Finally, it is widely-used for processing spatial data,
trips correspond to vehicles (which are subdivided into three which results in a sizable online community and support. The
weight classes) while about 1% of all the trips are pedestrian primary disadvantage of using PostgreSQL is the limitation
movements. In addition, Figure 3 shows that most trips pertain regarding parallel queries (i.e., the planner will not conduct
to fleet vehicles. In total, the raw GPS traces include 1.4 billion a parallel query if it involves any data writing). However,
waypoints which requires 112 GB of storage space. this limitation will likely be removed in future releases of
PostgreSQL.
3) Penetration rate: Because the trajectory data represents
only a subset of vehicles on the road, it is important to
roughly quantify the penetration rate (PR) of the analyzed trips.
Doing so may help indicate the extent to which the sample
is representative of overall traffic, and also provide insight
into the total number of vehicles traveling on road segments
between fixed traffic sensors. To perform rough PR estimates,
we compared GPS traces and data from 47 automatic traffic
Fig. 3: Summary of trip attributes for 6.4 million October trips:
recorder (ATR) stations in Maryland, which typically provide
geospatial, mode, provider type and vehicle weight classes.
hourly vehicle counts without differentiating between vehicle
types. The average hourly PRs at 47 locations are provided in
Figure 5, which indicates that average PRs at these 47 locations
vary from 0.85% to 5.52%, with a median of 1.86%. This
implies that observed trips capture one in every 54 vehicles.

Fig. 4: OpenStreetMap routing tool is applied to map match B. Methods

time-stamped sequence of latitude/longitude pairs (left) to most We employ an array of machine learning algorithms and
likely road-based routes (right). data visualization techniques to extract value from 20 million
GPS traces and effectively communicate our results with
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 6

with a library of suitable similarity measures are integrated

Average Hourly Penetration Rate [%]

5
[42]. We used V-Analytics to do quick exploratory analysis,
compare performance of different clustering algorithms, and
4
obtain high-quality visuals.
QGIS is an open-source GIS tool developed through the
3
Open Source Geospatial Foundation [43]. QGIS was used to
prepare the majority of maps and map-based animations in this
2 work. We found QGIS particularly useful due to its interface
with PostgreSQL for easy preparation and management of
1 large datasets, its interface with Python for programmatic
manipulation of maps and their appearance, and the large
Fig. 5: Penetration rates of recorded trips are estimated via online community that provides support and numerous plug-
comparison with over 224,000 hourly records from 47 ATR ins written in Python and C++.
stations. The average hourly PRs vary over these 47 locations
from 0.85% to 5.52%, with the median of 1.86%.
C. O-D matrices
As argued in the literature review, demand modeling and
transportation planning relies on estimating the number of trips
transportation agencies. Here we provide an overview of the that take place between specific locations [2]. To illustrate the
clustering algorithms used in the analysis, as well as software value of trajectory data in estimating demand, we map the
solutions that the authors found particularly useful in analyzing origins and destinations of the 20 million trips to geographic
and visualizing trajectory data. This discussion should provide areas of different sizes (i.e., traffic analysis zones, zip codes,
a brief guideline to transportation agencies that are beginning counties and states), and visually explore the corresponding O-
to analyze trajectory data. D matrices. While dense O-D matrices are somewhat difficult
1) Density-based clustering: DBSCAN (density-based spa- to visualize, those with fewer entries can be visually explored
tial clustering of applications with noise) is a widely-applied using open-source software Circos [44]. For example, Figure
clustering algorithm [36], which identifies each data point as a 6a depicts GPS trips between Maryland and other states, where
core point, border point, or outlier based on two input parame- the green ribbons denote trips originating in Maryland and
ters: ε and M IN P TS. ε is a radius parameter that defines the ε- ending in other states. This visual indicates that most trips
neighborhood N (ε) around each point, and M IN P TS represents originate and end in few neighboring states (i.e., Virginia,
the minimum number of data points in N (ε) required to form Pennsylvania), which are ordered clock-wise based on the total
a core point. Clusters are built around core points (which number of trips. It also shows that the number of trips going
represent high-density areas) by iteratively adding density- in and out of Maryland is balanced, which can be observed
connected points. DBSCAN does not require the number of by comparing the two outermost concentric circles that are
clusters as an input parameter, can easily find arbitrarily- of approximately same length and color pattern. Figure 6b
shaped clusters, is robust with respect to outliers (which are visualizes the subset of these trips that traverse Maryland, and
treated as noise and do not affect existing clusters), and is also indicates that a notable number of trips that originate and end
implemented in many libraries which facilitates its application. in a neighboring state (e.g., District of Columbia, Delaware,
However, one of the disadvantages is that it is very sensitive Virginia) still use the Maryland infrastructure. Figure 6c shows
to input parameters [37], where small changes to the radius a county-based O-D matrix for trips internal to MD and
and distance parameters can yield different clustering results. suggests that most trips originate and end within the same
In addition, the definition of distance should be carefully county, which is an expected result because the median trip
considered because it naturally affects the results (e.g., see [38] length is about 7 miles (Figure 1). As previously mentioned,
for a related discussion of similarity measures for trajectories). we can also map GPS trips to smaller areas (e.g., zip codes
This paper utilizes DBSCAN for constructing isochrones based and traffic analysis zones), and explore the corresponding O-D
on trajectory data. Finally, a related density-based clustering matrices via interactive applications (e.g., GIS, web).
algorithm OPTICS (ordering points to identify the clustering Since the analyzed GPS traces represent only a sample of
structure) [39] is used in other applications, as described later all vehicles on the road, the O-D matrices shown in Figure 6
in the paper. need to be scaled by appropriate expansion factor(s) to estimate
2) Software: V-Analytics (formerly Descartes and Com- actual traffic. A rough estimate of the total number of trips
monGIS) is a free visual data exploration and visual analytics between regions can be obtained by scaling O-D matrices in
software that facilitate exploration, analysis and modeling of Figure 6 by the factor of 54 (see Section III-A3), which is the
different kinds of spatio-temporal data: events, time series, approach that the authors of this paper will take to obtain
trajectories and situations. The system includes a variety of an aggregate trip table needed as an input for a statewide
interactive visualization techniques [40], supports necessary transportation planning model. However, one could improve on
transformations of spatio-temporal data [41] and integrates this by trying to derive custom expansion factors for different
a number of computational methods, adapted for analysis in O-D pairs, days of the week, and hours of the day. This
space and time. Particularly, tools for clustering trajectory data analysis could be further improved by deriving O-D matrices
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 7

for different types of vehicles (i.e., passenger cars vs. trucks); for points within 10 min travel time of the origin, with outlier
however, this would require determining PR of trajectory data points colored red, non-outlier points colored brown, and an
for different types of vehicles (see Figure 3), which would isochrone defined as a single conforming 2-D boundary of non-
be possible with traffic sensors that can differentiate vehicle outlier points. Note that, if the algorithm had not removed
types. Unfortunately, this is not the case with Maryland ATR the marked outliers, the resulting concave hull would have
stations. included these points too, suggesting inflated levels of mobility.
This procedure is repeated for 20, 30, and 40 minute trips, and
the concave hulls bounding the non-outlier points are plotted in
D. An O-D pair Figure 9b. The shape of the different concave hulls reflects the
Rather than considering an entire O-D matrix at various fact that mobility is greatest along the main highways, which
levels of granularity, it is sometimes useful to focus on a matches our intuition. Finally, as a validation of the outlined
specific O-D pair. To illustrate this, we consider trips between approach, we note that isochrones for heavy vehicles designed
Washington and Baltimore (Figure 7), and use GPS traces based on a traditional method (Figure 9d) show very similar
between this O-D pair to visually explore flow patterns, travel patterns to those observed in Figure 9b.
time variability and split rates amongst three major routes. Fig- It is worth noting that suggested approach for constructing
ure 7a shows the raw trajectories as well as aggregated trips for isochrones via density-based clustering/filtering of trajectory
days and links between neighboring polygons. Interestingly, data, yields a different measure of accessibility than isochrones
both beltways and I-95 (the middle road) show clear weekly calculated from travel times. The proposed isochrones would
patterns, whereas I-295 (East-most road) has stable load with encompass locations where many people have traveled to
no weekly patterns. Moreover, Figure 7b visualizes travel times within a specified time period, whereas the latter show lo-
between the Washington and Baltimore beltways broken down cations which people could reach in the same period of
by hour of day for weekday/weekend and day of week. On time. Accordingly, some relatively close but less-visited (per-
weekdays, the morning peak occurs for trips departing at 7-8 haps unsafe or unpopular) areas may be excluded from the
AM, while the afternoon peak is observed for trips departing trajectory-based isochrones, thus providing a different picture
at 4-5 and 5-6 PM. A very different travel pattern is observed of accessibility to various facilities (e.g., supermarkets, gas
on weekends, during which travel times are much steadier and stations). Another advantage of designing isochrones based on
also shorter than on weekdays. trajectory data is that it can be carried out without information
about the transportation network and historical travel times
along various road links.
E. Trip generators and isochrones While the proposed density-based clustering approach rep-
In addition to analyzing GPS traces between O-D pairs, it is resents an innovative application of trajectory data to quantify
instructive to consider origins and destinations separately. For mobility, DBSCAN’s results are very sensitive to the input
example, Figure 8 shows trip origins, which are spread over parameters, where the best input parameters depend heavily
the entire state of Maryland. While the sheer number of data on the size and specific distribution of the dataset (e.g., note
points obscures any patterns, creating and overlaying a simple different parameter values reported in Figure 9c and different
heat map representing origin density shows that many of the number of points mentioned in the previous paragraph). Conse-
trips originate at only a handful of locations. The main trip quently, the proposed approach suffers from excessive param-
generators are downtown Baltimore, Baltimore-Washington eter tuning and the need for visual sanity checks. In particular,
International Airport, and the stretch between Bethesda and parameter setting includes a trial and error approach with the
German Town. Upon identifying the main trip generators, we goal of having DBSCAN provide a large cluster of points
can query trips that originate in these areas and use their tra- around the origin location that neither encompasses remote
jectories to construct isochrones. However, trajectory datasets waypoints nor excludes areas with many waypoints (see Figure
often contain anomalous waypoints, which may skew mobility 9a for an example). As an extension of the proposed approach,
statistics and visualizations. Here we describe a density-based one could try to develop a method to automatically adjust
clustering approach that helps identify these outliers using the parameter setting for different case studies. Development of
previously-described DBSCAN algorithm. such a method would certainly represent a challenging task.
To showcase this approach, we consider a set of trips
originating from a single location and use the DBSCAN
algorithm to identify outliers for 10, 20, 30, and 40 minute F. Public transit
trips. As an example, we focus on a set of approximately 3,000 Public transit operates most efficiently when it provides ser-
trips beginning from the Port of Baltimore, which consists of vices that appropriately match customers’ spatial and temporal
218,302 total points (95,155 within 10 min, 141,586 within demand. Since GPS traces capture spatio-temporal patterns,
20 min, 164,053 within 30 min, and 178,257 within 40 min). they can be used to improve public transit by comparing
Using the scikit-learn Python implementation of DBSCAN existing transit routes with actual trips in a metropolitan
[45], we cluster the points for different combinations of input region. To illustrate this application, we focus on trips in the
parameters (Figure 9c), remove the points that algorithm Annapolis, MD region and cluster their O-D pairs using the
identifies as outliers, and visualize the results in the form of OPTICS algorithm [39]. The clustered O-D pairs are color-
isochrones. Figure 9a shows the results of running DBSCAN coded and shown in Figure 10a. The map-matched trajectories
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 8

AR

Somerse
WI
IA
WI

AR
IA

Dorc
MS

Carolin
MS

ore
Kent
FL

FL
AL

VT

m
heste
ME

ME

Calv

Balti
t
e
IL

Ga
IL

r
ert
VT

NH

Al

rre
Qu
NH I

leg
AL

tt
ee

an
VA
M

nA
M
G

y
A R

nn
GA I Ta

es
RI D lb
M Sa ot
IN IN int
M
SC KY ary
Wo s
KY MA rce
ste
MA r ery
SC om
ntg
TN Cha Mo
TN rles
CT CT
NC Wicom
OH ico
NY
NC
OH Carroll
NY

NJ
Cecil
NJ

n Princ
to eGeo
WV Was
hing rges

DE rd
o
H arf
PA
DE
VA

k
ric
e de

An
Fr

n
eA
ard

r
V

un
W

de
How

BaltimoreCity

l
DC

DC
PA

(a) Trips between MD and other states (b) Trips traversing MD (c) Trips within and between counties in MD

Fig. 6: O-D matrices visualized with Circos [44].

(a) 10-min isochrone & outliers (b) Port of Baltimore isochrones

Isochrone ε M IN P TS
(min) (km) (pts)
10 1.1 60
20 1.3 20
30 1.4 10
(a) Trips from Baltimore to Washington (b) Trip statistics
40 1.6 5

Fig. 7: Trajectories of trips between Washington and Baltimore (c) DBSCAN parameters (d) Traditional isochrones
beltways that took place during October. Boxplots show travel
times for trips between the two beltways. Fig. 9: DBSCAN with the outlined parameters is used to con-
struct isochrones from trip waypoints. After filtering waypoints
based on density, the isochrone is obtained by constructing a
concave hull which connects the boundary points (see Figure
9a for an example). Traditional isochrones for heavy-vehicles
from OpenRouteService [46] are used to validate the proposed
clustering-based approach (compare Figures 9b and 9d).

(a) Trip origins in MD (b) Heat map of origins in MD

are then overlaid onto the existing Annapolis transit network in
Figure 10b, applying a linear heat map in order to emphasize
Fig. 8: Some of the major trip generators: Baltimore down- the most-traveled routes. This visual comparison of important
town, BWI airport – Fort Mead, Bethesda – German Town. trajectories and the transit network reveals that some highly-
traveled routes are currently not covered with the transit
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 9

(a) Washington-Baltimore (b) I-95 nearby North Laurel

Fig. 11: Heat map of locations with higher speed recordings

(a) O-D clusters and their raw (b) Heat map of trajectories vs. indicates candidate locations for implementation of speed cam-
trajectories transit network eras. After an initial analysis at the regional level (Figure 11a),
an analyst can focus on a particular road segment and explore
Fig. 10: Clusters of trips in Annapolis can be used to modify directional speed profiles (Figure 11b). Color thresholds can
bus transit network in order to accommodate additional move- be changed to narrow down candidate locations.
ments. The last visual contrasts commonly traveled routes with
the transit system shown with solid black line.
truckers quickly learn the locations of these systems and can
start taking detours in order to avoid them, which can lead to
system. This simple visual comparison may be useful for facil- increased pavement and environmental damage due to more
itating discussion with the City of Annapolis about modifying vehicle miles traveled [47], [48].
bus routes to best accommodate additional trips. Furthermore, Trajectory data can reveal route choices of millions of
given sufficient interest in a full transit system evaluation or anonymized drivers, which can be used to investigate the extent
re-design, the GPS traces could be used in conjunction with an to which truckers are avoiding WIM systems. As an illustrative
array of data mining, operations research and microsimulation case study, we consider two WIM systems in Maryland and
techniques to explore spatio-temporal characteristics of trips, compute the percentage of vehicles that take immediate detours
optimize routes and service frequencies, and evaluate potential (Figure 12). The results indicate that trucks above 26k LB
savings. are not bypassing the systems, whereas 0.6% − 1.8% of other
vehicles are deviating from the main road in the immediate
vicinity of the WIM systems and then returning to the main
G. Safety road afterwards. This may suggest an evasion problem, because
Detailed trajectory data can reveal speed profiles of millions at least one third of these trips incurred greater travel times by
of anonymized drivers, which has important safety implica- taking detours. Also, in our data we are unable to differentiate
tions. We compute average speeds between all consecutive between passenger cars that would not have an incentive to
waypoints in our data set (which includes 1.4 billion GPS avoid WIM systems and trucks below 14k lb, so the percent
points), and focus on ones with higher than average speeds. of small trucks taking detours may be much higher. It is
Figure 11 shows a heat map that indicates locations where noteworthy that considering additional alternative routes would
higher speeds are recorded with greater frequency. The result provide a better picture of potential evasive strategies. Again,
could be readily used by agencies in charge of deploying speed we stress that trajectory data is anonymized and potential
cameras and radar patrols, which would likely help improve evasions cannot be traced back to individuals; the objective
safety and reduce property damage. However, we stress here is to identify areas that may be good candidates for additional
that trajectory data is anonymized and speeding cannot be weight control, which would reduce excessive damages and
traced back to individuals; the goal is to identify segments also improve safety for all the road users.
of the road network that may be good candidates for safety
improvements. IV. D ISCUSSION
Since characteristics of trajectory data can significantly
H. Weight control influence its applicability and thereby value, we provide a
Some truckers may overload their vehicles in order to in- discussion about some possible challenges that transportation
crease their productivity and profits, which results in excessive agencies should be aware of when purchasing trajectory data.
pavement and environmental damages. An effective way of The following is a list of potential data-related issues that
reducing this damage is to implement weigh-in-motion (WIM) agencies may want to discuss with data vendors in order to
systems, which are designed to detect and fine overweight obtain a more complete picture about applicability of a spe-
trucks. However, an issue with these systems is that they cific dataset. Some general recommendations to transportation
are inroad facilities, which once deployed in a transportation agencies interested in acquiring trajectory data are included as
network remain in their locations for several years. Thus, well.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 10

Main Route Weight Class Main Road Circumvent long-distance trips, agencies from multiple states could jointly
Detour (1,000 lb) (veh) (%) purchase data for an entire region (e.g., East Coast or all of
[0, 14] 3,794 1.45 USA), which also may be more cost efficient due to economies
(14, 26) 12,333 0.61
Main Route [26, ∞) 4,847 0.00
of scale.
Detour
[0, 14] 1,775 1.82 4) Population bias: Transportation agencies should be aware
(14, 26) 6,717 1.03 of the bias in data towards certain types of vehicles. For
[26, ∞) 6,403 0.02 example, the dataset discussed in this paper is biased towards
delivery trucks (Figure 3). This may not represent a major issue
Fig. 12: Examining potential evasion of WIM systems at if the observed region includes a network of ATR stations that
MD 32 East (left) and US-301 North (right) along immediate can differentiate between different vehicle types. In this case,
detours. Locations of WIM systems along the main routes are an analyst can determine the penetration rates of different types
indicated with pentagrams. of vehicles (passenger cars vs. trucks) and account for any bias
in further analysis. However, when such a network of sensors
is unavailable, correcting for the bias becomes a challenge and
may limit applications of trajectory data (e.g., estimation of an
1) Sampling rate: The average time lapse between consec- O-D matrix becomes a challenge). Therefore the government
utive waypoints significantly affects applicability of trajectory agencies interested in purchasing trajectory data should also
data, and agencies should try to acquire data with the highest account for the availability of other data sources that would
granularity possible (e.g., with the median or average time enable them to correct for the aforementioned bias in data.
lapse of 1 second). For example, a large time lapse between
5) Unique device identifications: Each trip in a trajectory
waypoints may not influence estimation of O-D matrices, but it
dataset includes an identification (ID) of the device it was
could make reconstruction of road-based trajectories a signif-
recorded from. Device IDs enable an analyst to chain con-
icant challenge, especially in dense urban areas where it may
secutive trips of the same vehicle and thereby reconstruct its
be impossible to determine which route a vehicle took. Thus,
movement over a longer period of time, which provides a better
it is important to request information about the granularity of
insight into mobility patterns. However, data vendors may
data and assess how it would influence the anticipated analysis
decide to periodically change device IDs (e.g., at midnight)
before actually acquiring data. Also, requesting road-based
for privacy or some other reasons, which clearly limits the
trajectories in addition to raw data, may save agencies quite
analysis. Thus, transportation agencies interested in purchasing
a bit of time and resources needed for map matching, which
trajectory data should inquire about vendor’s policies with
was discussed in Section III-A1.
respect to resetting device IDs and account for its implications
2) Spatial precision: The number of decimal numbers used on their analyses. Additional issues that analysts should be
to report waypoint latitudes/longitudes is another factor that aware of are occasionally duplicated or swapped device IDs,
can influence applicability of trajectory data. For example, which may arise when resetting device IDs. These and other
rounding a waypoint location to four decimal numbers intro- issues related to trajectory data are discussed in [49].
duces an error of about 11 m. While this error would not
necessarily prevent us from reconstructing road-based trajec-
tories or studying demand, it would significantly affect speed V. C ONCLUSIONS
estimates and its use in microsimulation models. Assuming
that the median spacing between two consecutive waypoints This paper synthesizes innovative applications of trajectory
is 28 m (Figure 2), location errors of 11 m would make data in road transportation, which is relevant to government
speed estimates meaningless. The same applies to computing agencies looking to introduce this type of data into their anal-
vehicle acceleration/deceleration rates that are needed for mi- yses and decision making processes. We provide a literature
crosimulation models used to estimate emissions, such as VT- review illustrating applications of trajectory data in six areas
Micro [23]. Thus, agencies should request latitudes/longitudes of road transportation systems analysis: demand estimation,
expressed with six decimal numbers, and still account for the modeling human behavior, designing public transit, traffic per-
errors that are inherent to GPS technology. formance measurement and prediction, environment and safety.
3) Division of trajectories into trips: Transportation agen- Additionally, we perform an extensive analysis of 20 million
cies should be aware that GPS companies may reset a trip GPS trajectories in Maryland, demonstrating both existing and
whenever the vehicle is idle for a specified period of time new applications of trajectory data in transportation. We em-
(e.g., 10 minutes). When this occurs within the boundaries ploy an array of techniques encompassing data processing and
of a state for which data was purchased, an analyst can management, machine learning, and visualization, and describe
still chain consecutive trips by looking at the unique device best-practices for using them to extract value from trajectory
identifications. However, when a trip gets reset once it leaves data, thus allowing transportation agencies to estimate the time
the state, than the information about subsequent lags of the and effort needed to introduce this type of data into their
trip is lost. This is probably the reason that Figure 2 does modeling efforts. As trajectory data becomes more prevalent
not include any trips going to the West Coast, as such a and acquisition costs decrease, we believe that this type of data
long trip would necessitate stops long enough to reset the will become an invaluable resource to transportation agencies
trip. To overcome this problem and gain better insight into across the world.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 11

ACKNOWLEDGMENT [16] T. M. Brennan Jr, S. M. Remias, and L. Manili, “Performance measures

to characterize corridor travel time delay based on probe vehicle
The authors would like to thank Subrat Mahapatra and data,” Transportation Research Record: Journal of the Transportation
the Maryland State Highway Administration for their support Research Board, no. 2526, pp. 39–50, 2015.
throughout this project. Help from the I-95 Corridor Coalition [17] D. Wedin, “Travel time estimation in Stockholm using historical GPS
and the City of Annapolis are also appreciated. The last data,” Master’s thesis, Uppsala University, Uppsala, Sweden, 6 2015.
two authors also acknowledge support by EU projects VaVeL [18] G. Evensen, Data assimilation: The ensemble Kalman filter. Springer
“Variety, Veracity, VaLue: Handling the Multiplicity of Urban Science & Business Media, 2009.
Sensors” (grant agreement 688380) and Track&Know “Big [19] W. Wei, G. Xiucheng, J. Jing, and R. Bin, “GPS probe based freeway
Data for Mobility Tracking Knowledge Extraction in Urban real-time travel speed estimation using Kalman filter,” in Intelligent Sys-
Areas” (grant agreement 780754). This support is gratefully tem Design and Engineering Application (ISDEA), 2010 International
Conference on, vol. 1. IEEE, 2010, pp. 797–800.
acknowledged, but it implies no endorsement of the findings.
[20] J.-S. Yang, “Travel time prediction using the GPS test vehicle and
Kalman filtering techniques,” in Proceedings of the 2005, American
R EFERENCES Control Conference, 2005. IEEE, 2005, pp. 2128–2133.
[21] EPA, “Inventory of U.S. greenhouse gas emissions and sinks: 1990-
[1] G. Andrienko, N. Andrienko, W. Chen, R. Maciejewski, and Y. Zhao, 2014,” Tech. Rep., 2016.
“Visual analytics of mobility and transportation: State of the art and
further research directions,” IEEE Transactions on Intelligent Trans- [22] J. M. Bandeira, T. Fontes, S. R. Pereira, P. Fernandes, A. Khattak,
portation Systems, 2017. and M. C. Coelho, “Assessing the importance of vehicle type for
the implementation of eco-routing systems,” Transportation Research
[2] M. S. Iqbal, C. F. Choudhury, P. Wang, and M. C. González, “Devel-
Procedia, vol. 3, pp. 800–809, 2014.
opment of origin–destination matrices using mobile phone call data,”
Transportation Research Part C: Emerging Technologies, vol. 40, pp. [23] H. Rakha, K. Ahn, and A. Trani, “Development of VT-Micro model
63–74, 2014. for estimating hot stabilized light duty vehicle and truck emissions,”
[3] J. L. Toole, S. Colak, B. Sturt, L. P. Alexander, A. Evsukoff, and Transportation Research Part D: Transport and Environment, vol. 9,
M. C. González, “The path most traveled: Travel demand estimation no. 1, pp. 49–74, 2004.
using big data resources,” Transportation Research Part C: Emerging [24] T. Feng, T. Arentze, and H. Timmermans, “Instantaneous emission
Technologies, vol. 58, pp. 162–177, 2015. modeling with GPS-based vehicle activity data: Results of diesel trucks
[4] Y. Zheng and X. Xie, “Learning travel recommendations from user- for one-day trips,” in Proceedings of the Eastern Asia Society for
generated GPS traces,” ACM Transactions on Intelligent Systems and Transportation Studies, vol. 2011, no. 0. Eastern Asia Society for
Technology (TIST), vol. 2, no. 1, p. 2, 2011. Transportation Studies, 2011, pp. 147–147.
[5] Y. Zheng, “Trajectory data mining: An overview,” ACM Transactions [25] J. Yang, J. Dong, and L. Hu, “A data-driven optimization-based
on Intelligent Systems and Technology (TIST), vol. 6, no. 3, p. 29, 2015. approach for siting and sizing of electric taxi charging stations,”
Transportation Research Part C: Emerging Technologies, vol. 77, pp.
[6] N. Andrienko, G. Andrienko, G. Fuchs, and P. Jankowski, “Scalable and
462–477, 2017.
privacy-respectful interactive discovery of place semantics from human
mobility traces,” Information Visualization, vol. 15, no. 2, pp. 117–153, [26] M. Berlingerio, B. Ghaddar, R. Guidotti, A. Pascale, and A. Sassi, “The
2016. graal of carpooling: Green and social optimization from crowd-sourced
[7] C. M. Schneider, V. Belik, T. Couronné, Z. Smoreda, and M. C. data,” Transportation Research Part C: Emerging Technologies, vol. 80,
González, “Unravelling daily human mobility motifs,” Journal of The pp. 20–36, 2017.
Royal Society Interface, vol. 10, no. 84, p. 20130246, 2013. [27] Y. Hara and M. Kuwahara, “Traffic monitoring immediately after a
[8] M. E. Ben-Akiva and S. R. Lerman, Discrete choice analysis: Theory major natural disaster as revealed by probe data – A case in Ishinomaki
and application to travel demand. MIT press, 1985, vol. 9. after the Great East Japan Earthquake,” Transportation Research Part
A: Policy and Practice, vol. 75, pp. 1–15, 2015.
[9] Y. Xu and M. C. González, “Collective benefits in traffic during mega
events via the use of information technologies,” 2016, unpublished. [28] X. Song, Q. Zhang, Y. Sekimoto, and R. Shibasaki, “Intelligent system
for urban emergency management during large-scale disaster,” in Pro-
[10] A. Lima, R. Stanojević, D. Papagiannaki, P. Rodriguez, and M. C.
ceedings of the Conference on Artificial Intelligence (AAAI14), 2014,
González, “Understanding individual routing behaviour,” Journal of The
pp. 458–464.
Royal Society Interface, vol. 13, no. 116, p. 20160021, 2016.
[11] F. Pinelli, R. Nair, F. Calabrese, M. Berlingerio, G. Di Lorenzo, and [29] Y. Ikeda and M. Inoue, “An evacuation route planning for safety route
M. L. Sbodio, “Data-driven transit network design from mobile phone guidance system after natural disaster using multi-objective genetic
trajectories,” IEEE Transactions on Intelligent Transportation Systems, algorithm,” Procedia Computer Science, vol. 96, pp. 1323–1331, 2016.
vol. 17, no. 6, pp. 1724–1733, 2016. [30] M. Dozza and J. Werneke, “Introducing naturalistic cycling data: What
[12] V. Guihaire and J.-K. Hao, “Transit network design and scheduling: A factors influence bicyclists safety in the real world?” Transportation
global review,” Transportation Research Part A: Policy and Practice, Research Part F: Traffic Psychology and Behaviour, vol. 24, pp. 83–
vol. 42, no. 10, pp. 1251–1273, 2008. 91, 2014.
[13] M. Berlingerio, F. Calabrese, G. Di Lorenzo, R. Nair, F. Pinelli, and [31] J. Strauss, L. F. Miranda-Moreno, and P. Morency, “Mapping cyclist
M. L. Sbodio, “AllAboard: A system for exploring urban mobility activity and injury risk in a network combining smartphone GPS data
and optimizing public transport using cellphone data,” in Joint Euro- and bicycle counts,” Accident Analysis & Prevention, vol. 83, pp. 132–
pean Conference on Machine Learning and Knowledge Discovery in 142, 2015.
Databases. Springer, 2013, pp. 663–666. [32] N. Koshak and A. Fouda, “Analyzing pedestrian movement in Mataf
[14] T. M. Brennan Jr, S. M. Remias, G. Grimmer, D. Horton, E. Cox, and using GPS and GIS to support space redesign,” in The 9th international
D. Bullock, “Probe vehicle-based statewide mobility performance mea- conference on design and decision support systems in architecture and
sures for decision makers,” Transportation Research Record: Journal urban planning, 2008.
of the Transportation Research Board, no. 2338, pp. 78–90, 2013. [33] A. Johansson and D. Helbing, “Analysis of empirical trajectory data of
[15] A. Kesting and M. Treiber, Traffic Flow Dynamics: Data, Models and pedestrians,” in Pedestrian and Evacuation Dynamics 2008. Springer,
Simulation. Springer: Berlin, 2013. 2010, pp. 203–214.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 12

[34] M. Haklay and P. Weber, “Openstreetmap: User-generated street maps,” Przemysław Sekuła received his Ph.D. degree in
IEEE Pervasive Computing, vol. 7, no. 4, pp. 12–18, 2008. management from the University of Economics in
[35] P. Newson and J. Krumm, “Hidden Markov map matching through noise Katowice in 2012. His research interests include
and sparseness,” in Proceedings of the 17th ACM SIGSPATIAL inter- applications of machine learning and artificial intelli-
national conference on advances in geographic information systems. gence in transportation. Currently, he is working as a
ACM, 2009, pp. 336–343. researcher at the Center for Advanced Transportation
Technology, University of Maryland, USA, and an
[36] M. Ester, H.-P. Kriegel, J. Sander, X. Xu et al., “A density-based Assistant Professor at the University of Economics
algorithm for discovering clusters in large spatial databases with noise,” in Katowice, Poland.
in Kdd, vol. 96, no. 34, 1996, pp. 226–231.
[37] G. Karypis, E.-H. Han, and V. Kumar, “Chameleon: Hierarchical
clustering using dynamic modeling,” Computer, vol. 32, no. 8, pp. 68–
75, 1999.
[38] N. Pelekis, G. Andrienko, N. Andrienko, I. Kopanakis, G. Marketos, and
Y. Theodoridis, “Visually exploring movement data via similarity-based
analysis,” Journal of Intelligent Information Systems, vol. 38, no. 2, pp.
343–391, 2012.
[39] M. Ankerst, M. M. Breunig, H.-P. Kriegel, and J. Sander, “OPTICS: Zachary Vander Laan received his M.S. degree
Ordering points to identify the clustering structure,” in ACM Sigmod in civil engineering from the University of Mary-
Record, vol. 28, no. 2. ACM, 1999, pp. 49–60. land in 2017. His research interests include intelli-
[40] N. Andrienko and G. Andrienko, Exploratory analysis of spatial and gent transportation systems, data visualization, and
temporal data: a systematic approach. Springer Science & Business applications of machine learning in transportation.
Media, 2006. Currently, he is working at the Center for Advanced
Transportation Technology, University of Maryland,
[41] G. Andrienko, N. Andrienko, P. Bak, D. Keim, and S. Wrobel, Visual USA.
analytics of movement. Springer Science & Business Media, 2013.
[42] G. Andrienko, N. Andrienko, S. Rinzivillo, M. Nanni, D. Pedreschi,
and F. Giannotti, “Interactive visual clustering of large collections of
trajectories,” in Visual Analytics Science and Technology, 2009. VAST
2009. IEEE Symposium on. IEEE, 2009, pp. 3–10.
[43] QGIS Development Team, QGIS Geographic Information System, Open
Source Geospatial Foundation, 2015.
[44] M. Krzywinski, J. Schein, I. Birol, J. Connors, R. Gascoyne, D. Hors-
man, S. J. Jones, and M. A. Marra, “Circos: An information aesthetic
for comparative genomics,” Genome research, vol. 19, no. 9, pp. 1639– Gennady Andrienko is a Lead Scientist responsible
1645, 2009. for the visual analytics research with Fraunhofer
Institute Intelligent Analysis and Information Sys-
[45] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, tems and a Professor (part-time) with City Univer-
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg et al., sity London. He has co-authored two monographs,
“Scikit-learn: Machine learning in Python,” Journal of Machine Learn- Exploratory Analysis of Spatial and Temporal Data
ing Research, vol. 12, no. Oct, pp. 2825–2830, 2011. (Springer, 2006) and Visual Analytics of Movement
[46] P. Neis and A. Zipf, “Openrouteservice. org is three times “open”: (2013), and more than 80 peer-reviewed journal
Combining opensource, openls and openstreetmaps,” GIS Research UK papers. From 2007 to 2015, he was chairing the ICA
(GISRUK 08). Manchester, 2008. Commission on GeoVisualization. He co-organized
[47] N. Marković, I. O. Ryzhov, and P. Schonfeld, “Evasive flow capture: scientific events on visual analytics, geovisualization,
Optimal location of weigh-in-motion systems, tollbooths, and security and visual data mining, and co-edited 13 special issues of journals.
checkpoints,” Networks, vol. 65, no. 1, pp. 22–42, 2015.
[48] ——, “Evasive flow capture: A multi-period stochastic facility location
problem with independent demand,” European Journal of Operational
Research, vol. 257, no. 2, pp. 687–703, 2017.
[49] G. Andrienko, N. Andrienko, and G. Fuchs, “Understanding movement
data quality,” Journal of location Based services, vol. 10, no. 1, pp.
31–46, 2016. Natalia Andrienko has been with GMD, currently
Fraunhofer Institute Intelligent Analysis and Infor-
mation Systems, since 1997. Since 2007, she has
been a Lead Scientist, where she has been involved in
visual analytics research. Since 2013, she has been a
Professor (part-time) with City University London.
She has co-authored the monographs Exploratory
Nikola Marković received his Ph.D. degree in trans- Analysis of Spatial and Temporal Data (Springer,
portation engineering from the University of Mary- 2006) and Visual Analytics of Movement (Springer,
land in 2013. His research interests include appli- 2013) and over 70 peer-reviewed journal papers. She
cations of operations research and machine learning received best paper awards at AGILE 2006, EuroVis
in transportation systems analysis. Currently, he is 2015, and IEEE VAST 2011 and 2012 conferences; best poster awards at
working at the Center for Advanced Transportation AGILE 2007, ACM GIS 2011, and IEEE VAST 2016; and VAST challenge
Technology, University of Maryland, USA. awards 2008 and 2014.

View publication stats