Comparative study on the effectiveness of various types of road traffic intensity detectors

Full text

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Comparative study on the effectiveness of various
types of road traffic intensity detectors
Andrzej Czyżewski, Sebastian Cygert,
Grzegorz Szwoch, Józef Kotus, Dawid Weber,
Maciej Szczodrak, Damian Koszewski
Multimedia Systems Department
Faculty of Electronics, Telecommunications and Informatics
Gdansk University of Technology
Gdansk, Poland
inznak@multimed.org
Andrzej Sroczyński, Tomasz Śmiałkowski
SILED Sp. z o.o.
andrzej.sroczynski@siled.pl, tomasz.smialkowski@siled.p
Kazimierz Jamroz, Wojciech Kustra
Highway and Transportation Engineering Department
Faculty of Civil and Environmental Engineering
Gdansk University of Technology
Gdansk, Poland
kjamroz@pg.edu.pl
wojciech.kustra@pg.edu.pl
Piotr Hoffmann
MICROSYSTEM Sp. z o.o.
p.hoffmann@microsystem.com.pl
Abstract— Vehicle detection and speed measurements are
crucial tasks in traffic monitoring systems. In this work, we
focus on several types of electronic sensors, operating on
different physical principles in order to compare their
effectiveness in real traffic conditions. Commercial solutions
are based on road tubes, microwave sensors, LiDARs and
video cameras. Distributed traffic monitoring systems require
a high number of monitoring stations. In order to improve the
accuracy of traffic monitoring, several modalities,
complementing each other, may be used in the monitoring
stations. In this paper, we propose a multimodal approach to
traffic monitoring, using sensors and signal processing
algorithms developed specifically for the described task. The
aim of the work described here is to test each modality in a
real-life scenario, assess their accuracy and to evaluate their
usefulness for multimodal traffic monitoring stations. The
modalities described in the paper are: Doppler sensor with
custom signal processing, video analysis based on cameras and
neural networks (employing deep learning algorithms), audio
monitoring based on an acoustic vector sensor developed by
the authors, as well as LiDAR and Bluetooth as supplementary
means of traffic monitoring. Additionally, road tubes and a
commercial video-based monitoring system were used in order
to provide reference data. Consequently, we are able to present
in this paper a comparative study on the effectiveness of traffic
sensors operating on the basis of different principles of work.
Keywords— traffic measurement; multimodal analysis;
signal processing
I. INTRODUCTION
The research presented in this paper is a part of a broader
project devoted to the development of intelligent road signs.
The scope of the whole project is covered by another paper
presented at the same conference, namely the paper entitled:
“Development of Intelligent Road Signs with V2X Interface
for Adaptive Traffic Controlling” authored by A. Czyżewski
et al. The following chapters present a part of the research
with focus to studying effectiveness of various traffic
intensity detectors. Sections II describes modalities,
algorithms and sensors that were used in the described work.
Section III presents results of the experiments in which
accuracy of each modality in traffic monitoring was
evaluated.
II. MODALITIES
A. Road tubes
Pneumatic tubes are one of the most common means of
measuring road traffic. A pair of rubber tubes is mounted in a
road, separated by some distance, is used to measure the
average speed of each vehicle within the measurement zone.
Additionally, axle counting is performed, so that a vehicle
class (given by its estimated width) may be determined. This
method is simple and accurate, but it is also invasive, as the
tubes obstruct the measured traffic. They are suitable only
for short-term measurements, they can be used on roads with
maximum two lanes in each direction, and they cannot be
used on road with high allowed speed (e.g. highways). They
also do not work accurately in case of traffic jams. However,
the road tubes may be used to collect the reference data, used
to assess performance of other sensors. Therefore, data from
the tubes were collected during the experiments described in
this paper and compared with results obtained with other
modalities.
B. Doppler radar
Doppler radars are the most common equipment used for
traffic speed monitoring and enforcement. They operate on
the basis of measuring the difference in frequency between
the transmitted signal and the signal reflected by a moving
object [1]. Professional radar equipment must comply to
strict requirements which results in their high cost. For the
purpose of collecting statistical data on road traffic in a
distributed system, a high number of low-cost sensors is
required. Therefore, we decided to employ consumer grade
motion detection sensors for this task. Such sensors are
characterized by a relatively low signal-to-noise ratio and are
susceptible to electromagnetic (EM) interference. The
proposed solution consists of a dual-channel (I/Q) motion
sensor, a signal amplifier, analog-to-digital converter and
Raspberry Pi microcomputer for signal analysis.
In order to make signal detection possible, noise and
interference has to be suppressed. The solution to this
problem developed by the authors is based on evaluating
difference between the phase spectra of signals recorded
from two I/Q output channels of the sensor, computed with
Fast Fourier Transform [#GS2]. For signals reflected by
moving objects, the phase difference should be about 90 or –
90 degrees, depending on the direction of movement, while
noise and EM interference should be concentrated around 0
degrees. Therefore, the phase difference is used to compute a
weighting function which is then multiplied by amplitude
spectra of the sensor signal. Thus, the level of noise and EM
interference is decreased, making further signal analysis
possible. Additionally, one of two directions (towards or
away from the sensor) may be eliminated, which allows for

separate detection of objects moving in each direction
(Fig. 1).
Fig. 1. Spectrograms of signal from the Doppler sensor (one channel):
original (upper plot) and processed with the proposed algorithm [2] (lower
plot) - only tracks of vehicles moving towards the sensor are retained. Plots
show spectral amplitude (color) vs. time (horizontal axis) and frequency in
Hz (vertical axis)
The remaining part of the algorithm is object detection
and tracking. New objects are detected when a new strong
spectral component is detected. This component is then
tracked in the successive signal sections. Since the sensor
measures only the radial component of the velocity vector,
frequency of the track decreases as the object moves towards
the sensor, also the track becomes wider, as the active length
of the object increases (Fig. 1). After a track is finished (an
object passed the sensor), the highest stable frequency of the
track is found and converted to velocity using the Doppler
equation [1]. A more detailed description of the algorithm
can be found in [2] and results of the experiments are
presented in Section IIIA.
The main purpose of the sensor described here is to
collect data on daily and hourly distribution of vehicle
number and traffic speed. It is not necessary to detect each
individual vehicle. There are several problematic cases in
which the Doppler radar performance is suboptimal. If there
are multiple lanes in the same direction, overlapping of
object tracks and object occlusion occur which results in
misdetections. If the vehicles move very close to each other,
the trailing vehicle is often undetected. Doppler radars are
also unable to detect traffic jams. Therefore, it seems
reasonable to supplement the Doppler detector with other
modalities.
C. Video analysis
Classical approach to vehicle counting in video is based
on using background subtraction methods [3]. Although
these methods are easy to deploy, energy-effective and they
do not require training, they may fail in difficult situations.
Vehicles might not be detected in traffic jam, shadows and
lights cause many false detections. Also, the background
subtraction method requires a stable camera view, being
sensitive to vibrations of the camera caused by the wind.
Current state-of-the art methods of object detection are
based on Convolutional Neural Networks (CNN) [4]. These
methods are extremely data hungry and they require
specialized processors to run. However, these requirements
are no longer a holdback for vehicle detection, since
specialized large datasets were published (i.e. UA-Detrac
[5]), and recently, energy-efficient neural networks
(SqueezeDet [6]) were developed. However, since
CNN-based algorithms are black-box approaches, it is not
known a priori what would be the accuracy of developed
model when deployed, especially when there are significant
differences between the source (train) and target (test)
domains. There are some methods that allow for
unsupervised learning, e.g. domain adaptation [7], however,
these methods are not explored in this paper. The goal of this
study is to measure how models trained on large datasets
performed in real traffic monitoring scenario.
A tracking algorithm is required in order to perform
vehicle counting. For this purpose, a simple online and real-
time tracking (SORT) [8] is used. It is a state-of-the art
tracking method that does not require training. In this
algorithm, a Kalman filter is used for the prediction step and
a Hungarian algorithm is used for associating new detections
with existing tracked objects. This method is very efficient
and accurate, its only drawback is the lack of handling long-
term occlusions, which is not an important problem in our
case.
Using video analysis for traffic monitoring presents
several advantages and disadvantages. Counting vehicles on
the basis of video analysis works also in multi-lane settings,
which has a great advantage over other modalities, making it
suitable for many real-world situations. While occlusions are
challenging, it is also possible for such an algorithm to work
in traffic jam situations. Video analysis can be also used for
other complementary tasks, such as anomaly detection.
However, these strengths come with some significant
drawbacks. Firstly, such algorithms tend to fail in
challenging illumination conditions, especially at night.
Further, accuracy of such a detector is dependent on the
position where the camera was placed, because it determines
the perspective with which the passing vehicles are observed.
This also means that accuracy on each of the lanes may be
slightly different. Finally, some manual calibration, such as
selecting region of interest where vehicles are counted, is
usually required.
The goal of this study if to measure usefulness of
CNN-based methods in traffic monitoring. For that purpose,
a CNN-based detector (SqueezeDet) was trained on
UA-Detrac dataset for vehicle detection. On top of that,
vehicle counting is performed and the results are compared
with Doppler radar and road tubes data. Also, 24h analysis is
performed in order to measure the detection accuracy, also
during the night. The results are presented in Section III.B.
D. Audio analysis
Analysis of audio signals from passing vehicles is
another method of traffic analysis. Compared with video
analysis, it is not susceptible to difficult conditions (e.g.
night). An Acoustic Vector Sensor (AVS) was used in the
research described in this paper. The AVS delivers the
following signals: acoustic pressure and three perpendicular
particle velocity components. These signals are used for the
calculation of three components of sound intensity on XYZ
axes [9]. Every passing vehicle produces noise during its
movement along the road, thus it can be considered to be a
moving sound source. Therefore, analysis of sound intensity
changes in time can be applied for counting vehicles and for
determining their direction of movement. The sound

intensity (SI) vector indicates the current position of the
sound source. The SI vector is updated 46 times per second,
thus it is possible to track the position of the sound source
(vehicle). Based on this data, presence of the vehicle and its
moving direction may be detected in a passive way, without
emitting any signals. This is the most important advantage of
this technique over active methods such as radars ad
LiDARs. In the presented research, the AVS based on
MEMS digital microphones was applied. The detailed
information about its design and properties can be found in
[9]. The proposed method was tested in real traffic
conditions. The obtained results are presented in Section
III.C.
E. LiDAR
Registration of vehicles in road traffic can also be carried
out using Light Detection and Ranging (LiDAR) devices
which emit a beam in the infrared spectrum. These signals
are series of very short light pulses which propagate in a
straight line, bounce off the vehicles and return to the
measuring device. Unlike radars, vehicle detecting via
LiDARs is less prone to measurement errors. This device
emits a more focused beam, which allows more accurate
measurement at a given point. LiDARs can be used both
during the day and at night, but they are more sensitive to
weather conditions than regular radars, these devices must be
also a static element of the system. In addition, the so-called
cosine error, which is the inaccuracy of the measurement
associated with the angle between LiDAR and the direction
of the vehicle movement, has to be considered. Because of
that, LiDARs are usually positioned alongside or
perpendicular to the road. The possible uses of LiDAR
include vehicle detection, lane occupancy and frequency of
vehicle flow
The main problem is that the beam is not directing
towards the passing car. This kind of measurement is
necessary to calculate speed of the vehicle. Theoretically,
two pulses of laser light allow the measurement of speed,
where the difference between two measurements is equal to
the distance divided by the time between two pulses is equal
to the speed of the car. In the case of such scenarios, it is not
possible to measure the speed due to the device’s position
perpendicular or at an angle in the direction of vehicle
movement, which results in inaccurate measurement and
calculation of the car speed is distorted by the cosine error.
The speed measurement obtained by the Doppler radar and
the occupancy time of the measurement point given by
LiDAR may be combined, allowing calculation of the
vehicle length. This data allows classification of the vehicle
and indication of belonging to particular groups. A series of
experiments were carried out and the results are presented in
section III.D.
F. Bluetooth
Employing Bluetooth technology for vehicle detection
becomes popular with growing number of vehicles equipped
with wireless communication systems. Depending on
Bluetooth device class, the detection range may vary
between 1 and 100 m [10]. Recent studies show that this
technology makes a promising method of collecting real-
time statistical traffic data and actual journey times from
measurements on a long distances, e.g. 1 km, which is not
possible or difficult with other modalities [11, 12]. The
general idea is based on recording anonymous Bluetooth
MAC addresses of devices together with a timestamp as they
pass by each detector and then perform matching the
addresses as vehicles pass through the next detector. Having
such data, it is possible to estimate average speed on a given
road section (e.g. a highway) and identify significant
differences from the expected results, like these caused by
congestions.
Technically, two approaches to MAC address monitoring
can be distinguished, namely basic and advanced. The basic
approach is utilizing standard hardware (Bluetooth adapter)
and scanning for nearby Bluetooth devices using the HCI
interface. The advanced approach requires a specialized
hardware: a radio module with software implementing the
Baseband controller and the Firmware Link Manager layers.
With these components, one can create a monitoring
application and optimize its temporal resolution. Following
the advanced approach, we constructed a practical vehicle
counter and performed field tests. The results of detecting
vehicles equipped with Bluetooth devices are presented in
Section III.E.
III. EXPERIMENTS AND RESULTS
In order to evaluate usefulness of the modalities
described in Section II, a number of field tests have been
carried out in real traffic conditions. The results of vehicle
counting (all modalities) and speed estimation (only the
Doppler device) are presented in this Section for all tested
sensors and compared with the reference data. The
experiments have been done at city streets located in the
vicinity of Gdańsk University of Technology. The reference
data was constructed from video recordings in which
vehicles were marked and counted by a human, and also
from the commercial device based on road tubes
(Metrocount MC5600 Vehicle Counter System). The
following traffic parameters were registered by this device
for each vehicle: time, speed, direction, traffic volume, axle-
based classification and gap between vehicles. The results
obtained for particular modalities are described in the
following subsections.
A. Doppler radar
The aim of the experiment was to compare performance
of the Doppler sensor (DS), supplemented with signal
processing algorithms developed by the authors, with the
commercial detector based on road tubes (RT). Both the
vehicle counting and the velocity measurement were
performed. The DS was mounted in a box placed on a
building wall at a height of 2.8 m, at the distance of 4.5 m
away from the road, aiming at the moving vehicles at a
horizontal angle of ca. 30 degrees. A total of 77 hours of
continuous recordings from both sensors (from 12th to 15th
October 2018) were analyzed. The data from RT were used
as a reference in order to assess the performance of the
custom detector based on DS (only vehicles moving towards
the sensor were analyzed). The results of vehicle counting
are summarized in Table I.
According to the obtained results, about 8% detections
made by DS were false (they were not detected by RT) and
about 3.8% detections made by RT were missed by DS. The
accuracy of DS in vehicle detection is therefore about 90%
which is sufficient for gathering statistical distribution of the
traffic in time, as shown in Fig. 2. However, not all FPs and
FNs may actually be the detection errors. The RT performs
the detection in a zone of the road (between two pairs of
tubes), while the DS detects vehicles at a point situated near

the entry to the RT’s detection zone. As a result, some
vehicles (e.g. those that parked within the detection zone)
may be detected by DS, but not by RT. It may explain the
relatively high ratio of false positives for the DS. Such cases
will be examined in more detail in the future work.
TABLE I. RESULTS OF VEHICLE COUNTING BY DOPPLER SENSOR
COMPARED WITH THE ROAD TUBES SYSTEM
Par.
Description
Total
Avg.
per
hour
ND
Number of vehicles counted
by the Doppler sensor (DS)
4371
56.8
NT
Number of vehicles counted
by road tubes (RT)
4182
54.3
TP
True positives
(detected by both sensors)
4022
(92% of ND,
96% of NT)
52.2
FP
False positives
(detected only by DS)
349
(8% of ND)
4.53
FN
False negatives
(detected only by RT)
160
(3.8% of NT)
2.08
Comparing the results of speed measurements, it was
observed that significant differences between both sensors
exist, with DS providing smaller values than RT in many
cases. The root mean squared difference was equal to 4.95
km/h. However, it should again be noted that RT performs a
zone measurement, while DS does a point measurement at
the beginning of the RT’s zone. It was observed that vehicles
do not move at a constant speed within the detection zone:
many of them accelerate coming out of a corner, some
decelerate approaching the crossroads while drivers are
looking for a parking space. It may be a reason for the
observed discrepancies in speed measurements, so the site
used for experiments is not optimal for assessing DS’s
accuracy in speed measurement (a zone where vehicles move
with approximately constant speed should be monitored,
instead).
Fig. 2. Hourly distribution of the detected number of vehicles for Doppler
sensor and for road tubes
B. Video analysis
For training of the classifier, the UA-Detrac dataset was
used. There are more than 140,000 frames in the
UA-DETRAC dataset, with 8250 vehicles that are manually
annotated, leading to a total of 1.21 million labeled bounding
boxes of objects. Data were collected by the dataset
providers from 24 different locations in China. The videos
are recorded at 25 fps, with resolution of 960×540 pixels.
A set of 60 videos was split into the training and the
validation set using 3:1 ratio. Training was stopped when the
f1-measure on validation set converged. Training and tests
were run on Intel Core i7, 2.6 GHz. Training on the
UA-Detrac dataset took 2 days. Input to the classifier was
resized to 480×270 pixels, as it provides a good compromise
between accuracy and detection time. Detection speed was
10.7 fps.
Regarding vehicle counting, a simple method of counting
vehicles that pass through the region of interest was
implemented. The specified region was manually selected
(green area in Fig. 3). For each bounding box, it was checked
whether its center lies within the region of interest. The
vehicle count is increased if the object identifier (returned by
the tracking component) was not counted before, and if the
trajectory length of the tracked vehicle is bigger than some
(small) threshold. This way, it is possible to ignore some of
the false positive detections caused mainly by low video
quality and shadows caused by the sunlight. When the
vehicle drives through the green area, the angle between its
trajectory and the road direction is measured. If this angle is
lower than 45 degrees, then the counter on the left side of the
road is increased, if is between 135 and 180 degrees, then the
counter on the right side of the road is incremented.
The video was recorded with 30 fps, however for fair
comparison only every 5th frame is used (which gives 6 fps),
as it is the estimated processing speed while the algorithm
was deployed to embedded system Nvidia Jetson TX2.
Fig. 3. View from camera and vehicle counting metrics
The goal of the experiment was to measure accuracy of
the vehicle counting system. The first challenge was the fact
that the analysis was performed during 24-hour period,
which includes nighttime detection known to be challenging
for video analysis. Interestingly, the proposed detector
achieved fair results also during nighttime, not a significant
drop in accuracy was noticed. This is because the UA-Detrac
dataset which was used for training contained few videos
recorded at nighttime and because the scene was moderately
enlightened by a street lamp. The second challenge came

from the fact that quality of the obtained video was very low,
mainly due to strong compression used while recording the
video. Gaussian smoothing was applied in order to remove
the artifacts, but the quality of the video was still very low.
Also, some dropouts were recorded in the video which
prevented a detection of some vehicles. Nevertheless, similar
technical challenges often occur in real world conditions.
The presented model achieved 87% recall and 92%
precision (Table II). Many of the false positives were caused
by traffic jam situations where the passing vehicles were not
counted by road tubes, but they still were mostly correctly
counted by video analysis. Some of false negatives were
caused by dropouts. It can be noted that strong sunlight and
shadows caused many false positives. Some vehicles were
recognized most accurately in the central part of the image,
which is probably caused by the fact that cars viewed from
such a perspective were frequent, i.e. cars recognition falls
down as the car is closer to the camera and the perspective is
vertical. This is one of the disadvantages of that system
which requires a careful selection of region of interest.
TABLE II. RESULTS OF VEHICLE COUNTING BY VIDEO ANALYSIS
COMPARED WITH ROAD TUBES
Par.
Description
Total
Avg.
per
hour
ND
Number of vehicles counted
by video analysis (VA)
1853
77.21
NT
Number of vehicles counted
by road tubes (RT)
1950
81.25
TP
True positives
(detected by both sensors)
1701
(92% of ND,
87% of NT)
70.88
FP
False positives
(detected only by VA)
152
(8% of ND)
6.33
FN
False negatives
(detected only by RT)
249
(3.8% of NT)
10.38
C. Audio analysis
It was not possible to perform the experiments with the
audio sensor at the same time as the previously described
ones, because the road tubes produced noise that would
distort signals recorded by the audio sensor. Therefore, a
separate experiment was performed on road with sparse
traffic flow. The AVS was placed at the distance of 2 m from
the road. An additional camera was used for collecting the
reference data. Detection of vehicles and their moving
direction was performed manually. The conditions were as
follows: temperature 11°C, pressure: 1015 hPa, wind speed
8.5 m/s, wind direction: North, road surface was dry. During
the observation period of 53 minutes, a total of 192 vehicles
passed the measurement point. Sound intensity signals were
recorded from the AVS, then an algorithm for vehicle
detection was applied. This algorithm analyzes sound
intensity values and time dependencies between the intensity
and the azimuth. The vehicle is detected when the sound
intensity level exceeds a defined threshold and the azimuth
crosses the zero line. Moreover, changes of the azimuth can
be used for determining the vehicle movement direction
(Fig. 4).
The algorithm was evaluated by comparing its results
with the reference data (Table #SI). Some false positive
results were observed (mostly on the further lane), no missed
detections occurred for the closer lane. Based on the
detection results, some metrics were calculated. Precision is
the number of true positives divided by the number of true
positives plus the number of false positives. Recall is the
number of true positives divided by the number of true
positives plus the number of false negatives. For the practical
application it is extremely important that the algorithm has a
high recall (low number of false negatives – no detection
while the expected event occurred). The presented algorithm
has a high recall and good accuracy. The obtained results are
very promising. The proposed method will be tested in a
longer period of time and under different volume of traffic
flow.
Fig. 4. Vehicle detection using the AVS: intensity vs. time (top) and
azimuth vs. time (bottom), with detected vehicles and their direction of
movement marked
TABLE III. RESULTS OF VEHICLE COUNTING USING AVS PROBE
Parameter
Lane 1
Lane 2
Total
True positive
98
92
190
False positive
6
9
15
False negative
0
2
2
Precision
94.2%
91.1%
92.7%
Recall
100.0%
97.9%
99.0%
Accuracy
94.2%
89.3%
91.8%
D. LiDAR
Two locations were selected for measurements. The first
one was a single-lane, two-way city street with a speed limit
of 50 km/h. The second location was a single-lane, two-way
village road with the speed limit of 90 km/h. Firstly, road
traffic measurements were made in three measurement
scenarios, both in day and night conditions. Measurements
were performed with the TF02 LIDAR device from
Benawake, with the measurement rate 100 Hz which allows
accurate readings of the passing vehicles, which is especially
important during the occupation of two lanes of the road. The
following experimental conditions were in force:
▪ measurement perpendicular to the road at the
distance of 100 cm from the roadway; the height
of the LIDAR fixation 40 cm from the surface,
▪ measurement of LIDAR set to 45-degree angle to
the road axis also at a height of 40 cm and at a
distance of 100 cm from the lane,
▪ measurement of LIDAR set to 45-degree to the
axis of the roadway and mounted at a height of 80
cm, and also at a distance of 100 cm from the
lane.
The TMS-SA4 device (a microwave radar for traffic
analysis) was used as a measurement reference, for counting
the number of vehicles and for measuring their speed.
According to the chosen reference, the effectiveness of each
scenario was evaluated and the optimal one was chosen. The

main collected data represented the number of detected cars.
In addition, the measurements were recorded with a
reference camera (Go Pro Hero 6 Black) and the vehicles
were counted manually. The results are shown in Table IV,
with false recognition rate (FRR) indicated.
TABLE IV. RESULTS OF THE MEASUREMENTS SCENARIOS
Scenario
Camera
(ref.)
TMS-SA
(ref.)
Correct
detections
FRR [%]
DAY
90deg/
40cm
48
48
46
4,17
NIGHT
90deg/
40cm
69
69
69
0
DAY
45deg/
40cm
43
43
42
2,32
NIGHT
45deg/
40cm
61
61
59
3,28
DAY
45deg/
80cm
45
45
40
11,11
NIGHT
45deg/
80cm
51
51
51
0
The results indicate the unambiguity of the choice of the
first scenario the for subsequent measurements. During the
second measurement session, 146 vehicles out of 151
identified by the camera (97%) were correctly detected by
the LiDAR. Setting the sensor perpendicular to the axis of
the road made it impossible to count vehicles in case of
occlusion (vehicles present on both lanes simultaneously),
which caused 5 missed detections (3%). No false positive
detections were observed.
E. Bluetooth
The evaluation of the Bluetooth vehicle detector was
done by comparing time instances of events detected by the
Bluetooth sensor with the reference data from the road tubes.
The maximum allowed time gap between the Bluetooth
detection and the vehicle counter detection was 5 seconds.
Hourly distribution of number of detected Bluetooth devices
and vehicle count is presented in Fig. 5 and Fig. 6.
Fig. 5. Hourly distribution of number of vehicles and detected Bluetooth
devices for weekday.
Fig. 6. Hourly distribution of number of vehicles and detected Bluetooth
devices for weekend
The characteristic of the measurement localization
implied careful Bluetooth data analysis. After data
preprocessing (i.e. rejecting extremely weak signals), the
total number of detected Bluetooth devices and vehicles was
equal to 81. Complete results regarding time of day are
provided in Table V, and the differences in observed mean
RSSI value according to the lane are shown in Table VI.
TABLE V. RESULTS OF BLUETOOTH DEVICE DETECTION IN DAYTIME
AND IN NIGHT TIME
Days (7-22)
Nights (22-7)
Positive match
57
7
Total devices
76
7
TABLE VI. MEAN RSSI OF BLUETOOTH A/V DEVICES DETECTED IN
CARS PASSED BY
Lane
mean RSSI [dB]
near
-82.5
far
-85.4
In the discussed scenario and localization, some
drawbacks concerning Bluetooth should be pointed out. The
first one is the case when a car maneuvers in the proximity of
the Bluetooth sensor, and it is not detected by the tubes.
Second, some strong signal emitting devices in vehicles
traveling on nearby streets may be detected. Moreover,
pedestrian walkway was along the road, so their hand-held
devices such as Bluetooth headphones might introduce
additional false detections.
IV. SUMMARY AND CONCLUSIONS
At the current stage of advancement of the project, it was
not possible to compare the data obtained from all types of
sensors, simultaneously. However, achieving this possibility
is expected in the next period. So far, the main focus has
been on assessing the effectiveness of individual sensors by
comparing the received data with the ground truth data and
on optimizing the operation principle and settings of
individual sensors. However, it is now already possible to
assess the effectiveness of individual sensors that work on
the basis of different physical principles. It seems that
microwave sensors and acoustic sensors have the best

application prospects for measuring traffic. For above
reasons the modality employing LiDARs and Bluetooth
should be treated as a complimentary functionality to be
exploited in the presence of others, more reliable ones.
Undoubtedly, the modality worth mentioning is video
analytics supported by the process of deep training of neural
networks. Despite the difficulties associated with proper
lighting of the scene, possible soiling of camera lenses and
the influence of bad weather, it brings interesting results,
especially in case where multiple lanes should be monitored,
simultaneously.
ACKNOWLEDGMENT
Project financed by the Polish National Centre for
Research and Development (NCBR) from the European
Regional Development Fund under the Operational
Programme Innovative Economy No. POIR.04.01.04-
0089/16 entitled: INZNAK – “Intelligent road signs with
V2X interface for adaptive traffic controlling”
REFERENCES
[1] G. Brooker, “Doppler Measurement,” In Sensors and Signals,
Australian Centre for Field Robotics, University of Sidney, 2006, pp.
443-462.
[2] G. Szwoch: “Suppression of distortions in signals received from
Doppler sensor for vehicle speed measurement,” 2018 Signal
Processing: Algorithms, Architectures, Arrangements, and
Applications (SPA), Poznań, 2018, pp.16-21.
[3] A. Bas, M. Tekalp, and F. S. Salman, “Automatic Vehicle Counting
from Video for Traffic Flow Analysis," 2007 IEEE Intelligent
Vehicles Symposium, Istanbul, 2007, pp. 392-397.
[4] S. Agarwal., J.O.D. Terrail, and F. Jurie, “Recent Advances in Object
Detection in the Age of Deep Convolutional Neural Networks,”.
arXiv preprint arXiv:1809.03193, 2018.
[5] S. Lyu et al., “UA-DETRAC 2017: Report of AVSS2017 & IWT4S
Challenge on Advanced Traffic Monitoring,” 2017 14th IEEE
International Conference on Advanced Video and Signal Based
Surveillance (AVSS), Lecce, 2017, pp. 1-7.
[6] B. Wu, F. Iandola, P. H. Jin, and K. Keutzer, “SqueezeDet: Unified,
Small, Low Power Fully Convolutional Neural Networks for Real-
Time Object Detection for Autonomous Driving,” 2017 IEEE
Conference on Computer Vision and Pattern Recognition Workshops
(CVPRW), Honolulu, HI, 2017, pp. 446-454.
[7] S. Cygert, and A. Czyżewski, “Vehicle detector training with labels
derived from background subtraction algorithms in video
surveillance,” 2018 Signal Processing: Algorithms, Architectures,
Arrangements, and Applications (SPA), Poznań, 2018, pp. 98-103.
[8] A. Bewley, et al., “Simple online and realtime tracking,” 2016 IEEE
International Conference on Image Processing (ICIP), 2016, pp.
3464-3468.
[9] J. Kotus, and G. Szwoch, “Calibration of acoustic vector sensor based
on MEMS microphones for DOA estimation,” Applied Acoustics 141
,2018, pp. 307–321, DOI: j.apacoust.2018.07.025.
[10] R. Friesen, and R. D. McLeod, “Bluetooth in intelligent transportation
systems: A survey,” International Journal of Intelligent Transportation
Systems Research, vol. 13, no. 3, pp. 143–153, Sep 2015, doi:
10.1007/s13177-014-0092-1.
[11] A. Haghani, M. Hamedi, K. F. Sadabadi, S. Young, and P. Tarnoff,
“Data collection of freeway travel time ground truth with bluetooth
sensors,” Transportation Research Record, vol. 2160, no. 1, pp. 60–
68, 2010, doi: 10.3141/2160-07.
[12] E. Sharifi, M. Hamedi, A. Haghani, and H. Sadrsadat, “Analysis of
vehicle detection rate for bluetooth traffic sensors: a case study in
Maryland and Delaware,” in Proceedings of the 18th world congress
on intelligent transport systems, 2011, pp. 16–20.