Motion Zone Estimation for Adaptive Radar-System

Parameters Selection
Galyna Kriukova Andrii Malenko Oleksandr Sudakov Sergii Vodopyan
National University Glushkov Institute of Cybernetics National Taras Shevchenko National University
of Kyiv-Mohyla Academy NAS of Ukraine University of Kyiv of Kyiv-Mohyla Academy
Kyiv, Ukraine Kyiv, Ukraine Kyiv, Ukraine Kyiv, Ukraine
kriukovagv@[Link] email address or ORCID email address or ORCID vodopyan.s.v@[Link]

Yevhenii Volynets
National University
of Kyiv-Mohyla Academy
Kyiv, Ukraine
email address or ORCID

Abstract—Recent developments in non-invasive radar mea- accuracy, making this band a good alternative for radar-sensing
surement of small-scale motion offer great potential for in- applications in worldwide industrial setting [?]. Rich point-
door and outdoor motion detection and activity classification cloud radar data (in the x, y and z-axes and radial velocity data)
in applications such as health diagnostics, health monitoring,
surveillance, occupancy sensing, and automotive. Combined with enables motion and object detection, localisation and classifi-
video technology, the measurement and recognition of motion cation in both indoor and outdoor applications. However, in
patterns will shape the future of home and workplace security, multi-target detection, typical FMCW radars will suffer from
safety, and comfort. This paper focuses on Doppler radar-based the range-velocity ambiguity problem, which causes ghost
sensing systems for motion monitoring. targets and missed targets. Gathering high accuracy meaning-
Index Terms—radio frequency sensor, noise reduction
ful data requires corresponding range and velocity maximum
I. I NTRODUCTION and resolution from the sensor. In order to optimise contrary
parameters by tuning chirp configuration, it is beneficial to
Motion detectors have found wide use in commercial
estimate and match actual field of view and motion zone of
applications, e.g. surveillance video cameras with infrared
sensor, therefore arises problem of position restoring: find
night vision, passive infrared sensors (PIR), thermal cam-
height and tilt angle of radar mounting according to given
eras, and radar sensors [?], [?], [?]. Recently, the radar-
point-cloud radar data.
based sensors have been employed in various active safety
applications and robotics, as low power and cost effective In this paper we propose a new 3D reconstruction method of
devices for motion detection, able to track and classify a motion zone for radar-based system for corresponding tuning
number of objects simultaneously and estimate corresponding and optimisation of system parameters for better performance.
distances and shapes. The Frequency-Modulated Continuous- We integrate domain knowledge to data science problem of
Wave (FMCW) technique, which is known for high resolution parameters optimisation, at the same time incorporate theo-
measurements, has been widely used in area of automotive retical models to analyse data, i.e. apply theory-guided data
radars, instrumentation and measurement. The basic idea of science approach discussed in [?]. Machine learning methods
FMCW radars is to attain the range and velocity information are widely used is signal processing, e.g. for noise reduction
from the beat frequency, which is composed of propagation or anomaly detection, in particular for UWB radar signal data
delay and Doppler frequency. [?], [?], [?], [?].
Most radio frequency sensors use the 24-, 60-, and 77- The organisation of the paper is as follows: background in-
GHz radio bands [?]. Spectrum regulations and standards formation and recent developments, followed by the theory of
developed by the European Telecommunications Standards In- Doppler radar sensing in noise reduction, motion detection and
stitute (ETSI) and U.S. Federal Communications Commission classification, experimental results for potential applications,
(FCC) restrict new products from using the 24-GHz ultra- and related challenges.
wide band and 77-GHz band for industrial factory, building,
and city infrastructure applications, including those requir- II. S YSTEM AND M ETHOD D ESCRIPTION
ing human/machine interaction (Electronic Communications
The radar-based system can be divided into two main sub-
Committee Decision (04)10). Sensors using 60-GHz band
systems: the hardware sub-system and software sub-system.
are capable of gathering rich point-cloud data with high
The hardware sub-system includes the RF sub-system (trans-
Identify applicable funding agency here. If none, delete this. mitter, receiver and antenna array) and Digital Signal Pro-
cessing (DSP) sub-system. The software subsystem consists where IFmax is maximum IF bandwidth supported, c is speed
of control part and algorithm part. of light, S is slope of the transmitted chirp.
In many applications it is important to be able to resolve
A. Hardware sub-system two closely spaced objects as two separate objects, rather than
The hardware configuration of the proposed radar system detect them as one. The smallest distance between two objects
includes the RF/analog sub-system and DSP sub-system. The that allows them to be detected as separate objects is referred
antenna array is controlled by DSP sub-system for 3 trans- to as range resolution. This primarily depends on the chirp
mitting antenna elements and 4 receiving antenna elements. A sweep bandwidth that the radar sensor can provide. The larger
commercial radar chip IWR6843 based on FMCW radar tech- the sweep bandwidth, the better the range resolution. TI’s radar
nology capable of operation in the 60-GHz to 64-GHz band devices support a 4 GHz sweep bandwidth that allows a range
is used in this system, which is made by Texas Instruments resolution of as low as approximately 4 cm.
[?]. Functional diagram of the sensor is presented in Fig. 1. c
Rangeresolution = , (2)
2B
where c is speed of light, and B is sweep bandwidth of FMCW
chirp.
As sweep bandwidth B of FMCW chirp is proportional
to its slope S, combining (1) and (2) we have that range
resolution and maximum of detection range are proportional
Rangeresolution ∝ Rangemax . Therefore, to optimise range
resolution, its better to minimise Rangemax to actual motion
zone.
Along with the distance, the relative velocity of the object is
another critical parameter of interest. The maximum measur-
able velocity in FMCW radars depends on the chirp cycle time,
that is, the time difference between the start of two consecutive
chirps. This in turn depends on how fast the frequency sweep
Fig. 1. IWR6843 Functional Block Diagram [?]. can be performed and the minimum inter-chirp time allowed.
λ
V elocitymax = , (3)
B. Software sub-system 4Tc
In this section we explain our adaptive method in details. where Tc is total chirp time, which includes chirp time and
The software subsystem has hardware control parts and algo- idle time, λ is wavelength of the signal used.
rithmic part. The software sub-system is based on firmware In applications one might need to separate out objects with
and millimeter wave software development kit (mmWave small velocity differences, for which good velocity resolution
SDK) provided by Texas Instruments [?]. It allows user to is needed. Velocity resolution mostly depends on the transmit
specify the chirping profile and displays the detected objects frame duration, that is, increasing the number of chirps in a
and other information in real-time. frame improves the velocity resolution.
λ
C. Chirp configuration V elocityresolution = , (4)
2N Tc
In linear FMCW radars, the transmit (TX) signal is a single where N is number of chiprs in a frame, Tc is total chirp time,
tone with its frequency changing linearly with time. This λ is wavelength of the signal used.
sweep in frequency is commonly referred to as a “chirp”. A Once again, combining (3) and (4) we get
set of these chirps form a “Frame” and this can be used as V elocityresolution ∝ V elocitymax .
the observation window for the radar processing. The various Doppler radars are generally ambiguous in either range or
parameters of the chirp ramp (like frequency slope, sweep doppler, or both. In case, then actual object velocity is more,
bandwidth, and so forth) impact the system performance [?]. than unambiguous V elocitymax various distortion and ghost-
The maximum and minimum distance over which a radar targets may be observed. In Fig. 2, 3 and 4 we compare
sensor can detect objects is an important parameter for a radar point-clouds of detected object with sensor with different chirp
sensor. Also, the range resolution (ability to distinguish two configs.
nearby objects) is another important metric. Detecting a far- It’s easy to see, that sensor performs much better, when
off object can be limited by either the SNR of the received V elocitymax exceeds object speed. Moreover, otherwise in-
signal or the intermediate frequency (IF) bandwidth supported ference in target localisation, ghost-targets and other distor-
by the radar device. The max range relationship with the IF tions occur. Thereby, as other configuration parameters are
bandwidth is shown in (1). interconnected and influence system parameters, adaptive chirp
IFmax · c configuration tuning is of interest for better performance in the
Rangemax = , (1) real-world applications.
2S
 Radial velocity of the object may be seen as a derivative of
the distance to object with respect to time:
dR
v= .
dt
Therefore, integrating it we should get
Z
R = v dt,
or Z t2
∆R = v dt. (5)
Fig. 2. Comparison of detected points of approaching object with speed t1
approx. 4.2 m/s, for chirp with V elocitymax = 15.3 m/s (left) and
V elocitymax = 3.4 m/s (right). Points colour corresponds to detected radial Therefore, having trajectory {(ti , xi , yi , zi , vi )| i =
velocity in m/s. 0, . . . , n} we may calculate approximation of the integral (5)
Z tk k
X
r̂k = v dt ≈ (ti − ti−1 )vi ,
t0 i=1

and compare the value r̂k with actual distance to object

q
rk = x2k + yk2 + zk2 .
For concerted trajectories values r̂k and rk have to agree
with each other up to some constant, corresponding to object
distance at initial time t0 . To verify it, it’s enough just to
consider linear regression for inputs {r̂k |k = 1, . . . , n} and
Fig. 3. Comparison of detected points of object moving tangentially with outputs {rk |k = 1, . . . , n}
speed approx. 13.9 m/s, for chirp with V elocitymax = 15.3 m/s (left) and
V elocitymax = 3.4 m/s (right). Points colour corresponds to detected radial r = ar̂ + b,
velocity in m/s.
where ideally a should be equal to 1 and b correspond to
distance to object at time t0 . Nevertheless, we may estimate
the coefficient of determination R2 of the prediction. The best
possible score is 1, in our numerical experiment we’ve used
different values of threshold to label trajectory as “pure” or
“noisy” one. At the same time, estimated value b, which ideally
corresponds to object distance at initial time, has to fit to
feasible distance, therefore, an inequality
0 ≤ b ≤ Rangemax + ε
must hold for some acceptable level of noise ε, for example
Fig. 4. Comparison of detected points of object with altering direction and
speed, for chirp with V elocitymax = 15.3 m/s (left) and V elocitymax = 0 ≤ b ≤ Rangemax + Rangeresolution .
3.4 m/s (right). Points colour corresponds to detected radial velocity in m/s.
This approach allows us to select trajectories corresponding
to salient motion, remove noisy and recurrent movement (e.g.
D. Noise reduction fan vibration or plant oscillation). Thereby the method enables
tuning chirp parameters according to estimated motion zone.
In real situations, a radar sensor produces imaginary targets
(the so-called ghost targets). These have nothing in common III. E XPERIMENT R ESULTS
with real objects, like vehicles or pedestrians, but are from Our experiments are conducted with a commercial FMCW
multi-path propagation of a transmitted radar wave or due to 60-64-GHz band radar chip IWR6843 made by Texas Instru-
interference from other radar sensors. ments [?]. The experiment is performed in a natural environ-
To eliminate noisy points from records, we apply the ment for various system mounting. Initial system parameters
following idea: extended Kalman filter (EKF) based multi- are the following according to selected chirp configuration:
hypothesis tracker is used to group points into trajectories. range resolution 0.28m, velocity resolution 0.25m/s, maximum
As the tracker doesn’t use information about estimated radial velocity 15m/s, maximum range 25m, angle resolution is up
object velocity, we use this information to detect noisy erro- to 5◦ on the sides of sensor’s field of view.
neous trajectories and remove corresponding points from data In order to verify motion zone reconstruction algorithm
for further analysis. performance, radar sensor is mounted at the height of 1.5m,
2m and 2.5m and tilt angle 0◦ , 15◦ , 30◦ and 45◦ . We record
identical sandglass-shaped trajectories of person moving in
front of radar sensor, and then we use our method to recon-
struct shape or corresponding point clouds. To measure the
performance of the algorithm, we estimate the plane underly-
ing the point cloud, compare its position (height and tilt angle)
with corresponding mounting height and tilt angle. Getting the
corresponding plane may be reduced to the total least squares
problem and solved by means of singular value decomposition
(SVD) method, i.e. for set of points we calculate centroid
(or centre of mass), subtract it from each point, compute
SVD decomposition, eigenvector corresponding to the least
eigenvalue defines plane as a normal to the plane vector.
In Fig. ??, ??, ??, ?? the noise removal and motion zone
reconstruction process is illustrated for tilt angles 0◦ , 15◦ , 30◦
and 45◦ , and for height 1.5m, 2.0m and 2.5m accordingly.

Fig. 7. Motion zone reconstruction for tilt angle 30◦ .

Fig. 5. Motion zone reconstruction for tilt angle 0◦ .

Fig. 8. Motion zone reconstruction for tilt angle 45◦ .

Two left-most plots depict point-cloud (view from above,

i.e. (x, y)-plane, and side view, i.e. (y, z)-plane, correspond-
ingly).
Two right-most plots correspond to the de-noised point-
cloud by means of tracker in the same projections. Points cor-
responding to static trajectories (recurrent motion) are pictured
with red colour. In (y, z)-plane green dashed line corresponds
to ground truth of footing level, blue one corresponds to
estimated plane of motion zone (approximately at the height
1.0m above the ground).
Results of motions zone reconstruction are presented in
Table ??. As we can see, the error of angle recovery is less than
10◦ , approximately 2◦ , which corresponds to measurement
error and/or angle resolution. Difference between estimated
middle plane of point-cloud and ground level definitely should
Fig. 6. Motion zone reconstruction for tilt angle 15◦ . be positive. As we can see, it’s approximately 1m, which
corresponds to estimated mass centre of the human body.
Evidently, for each mounting height and tilt angle we may
observe different point of the human body, and corresponding
middle plane of point-cloud may shift. Therefore, we assume,
that the angle error is more appropriate measure for estimating
performance of motion zone reconstruction.

TABLE I
M OTION ZONE RECONSTRUCTION

Ground Truth Estimated Difference

Figure angle height angle height angle height
Fig. ?? 0◦ 1.5m 0.7◦ 0.62 0.7◦ 0.88
Fig. ?? 0◦ 2.0m 2.6◦ 0.98 2.6◦ 1.02
Fig. ?? 0◦ 2.5m 3.2◦ 1.46 3.2◦ 1.04
Fig. ?? 15◦ 1.5m 14.0 ◦ 0.51 1.0◦ 0.99
Fig. ?? 15◦ 2.0m 14.5◦ 0.90 0.5◦ 1.1
Fig. ?? 15◦ 2.5m 13.1◦ 1.33 1.9◦ 1.17
Fig. ?? 30◦ 1.5m 32.2◦ 0.79 2.2◦ 0.71
Fig. ?? 30◦ 2.0m 31.1 ◦ 1.13 1.1◦ 0.87
Fig. ?? 30◦ 2.5m 32.5◦ 1.63 2.5◦ 0.87
Fig. ?? 45◦ 1.5m 48.6◦ 0.83 3.6◦ 0.67
Fig. ?? 45◦ 2.0m 54.8◦ 1.64 9.8◦ 0.36
Fig. ?? 45◦ 2.5m 43.4 ◦ 1.56 1.6◦ 0.94

According to estimated motion zone, the maximum range

doesn’t need to exceed 15m, therefore value of 25m is ex-
cessive, accordingly, chirp configuration may be modified to
decrease maximum detection range and improve range resolu-
tion up to 0.17cm. Moreover, having long history of records,
we may get estimation of maximum velocity observed, and
modify corresponding parameters accordingly.
IV. C ONCLUSIONS
In the paper we consider a new method for de-noising
and static clutter removal for motion zone reconstruction.
This approach may be used for adaptive change of chirp
configuration parameters for optimising radar system’s field
of view and performance.
ACKNOWLEDGMENT
Galyna Kriukova would like to express deep and sincere
gratitude to Mr. Sergii Bodnarchuk for providing invaluable
guidance and support throughout this research.
R EFERENCES