Radio-based Traffic Flow Detection and Vehicle

Classification for Future Smart Cities

Marcus Haferkamp1 , Manar Al-Askary1 , Dennis Dorn2 , Benjamin Sliwa1 , Lars Habel3 ,
Michael Schreckenberg3 and Christian Wietfeld1
1
Communication Networks Institute, TU Dortmund University, 44227 Dortmund, Germany
e-mail: {Marcus.Haferkamp, Manar.Askary, Benjamin.Sliwa, Christian.Wietfeld}@tu-dortmund.de
2
Wilhelm Schröder GmbH, 58849 Herscheid, Germany
e-mail: {Dorn}@mfds.eu
3
Physics of Transport and Traffic, University Duisburg-Essen, Germany
e-mail: {Habel, Schreckenberg}@ptt.uni-due.de

Abstract—Intelligent Transportation Systems (ITSs) providing accuracy are discussed, which compare the suitability of
vehicle-related statistical data are one of the key components for different machine learning algorithms and features for the
future smart cities. In this context, knowledge about the current defined problem statement. Moreover, the impact of ground-
traffic flow is used for travel time reduction and proactive jam
avoidance by intelligent traffic control mechanisms. In addition, reflected radio waves regarding the Received Signal Strength
the monitoring and classification of vehicles can be used in the Indicator (RSSI) is analyzed. Finally, the results show the high
field of smart parking systems. The required data is measured efficiency of the proposed approach and its suitability for being
using networks with a wide range of sensors. Nevertheless, in used in future smart cities.
the context of smart cities no existing solution for traffic flow
detection and vehicle classification is able to guarantee high
classification accuracy, low deployment and maintenance costs,
II. R ELATED W ORK
low power consumption and a weather-independent operation Traffic flow detection has been a topic of scientific interest
while respecting privacy. In this paper, we propose a radio-
based approach for traffic flow detection and vehicle classification
for a long time. In [2] a comparison of several approaches
using signal attenuation measurements and machine learning is presented. In recent years, the data of an increasing
algorithms. The results of comprehensive measurements in the number of various detection systems is aggregated building
field prove its high classification success rate of about 99%. a multi-functional data-driven ITS [3]. For this purpose, a
I. I NTRODUCTION wide range of different sensor and detection techniques with
specific advantages and disadvantages is used. A widespread
Intelligent traffic control mechanisms aim to reduce traffic
approach for vehicle detection and classification are camera-
jam occurrences and travel times as well as the 𝐶𝑂2 output.
based systems, which achieve a high classification success
These objectives are also key components for future smart
rate. Normally, several cameras are needed in those systems to
cities [1]. In order to achieve these goals, knowledge about the
analyze the scenario from different angles and perspectives. In
current traffic flow needs to be obtained at chosen measure-
contrast, an enhanced visual system which is able to categorize
ment locations. Apart from the intelligent traffic control, there
vehicles into various vehicle classes using a single camera is
are further application fields which can benefit from traffic
presented in [4]. Regardless of the lower number of cameras,
flow monitoring. For example, smart parking or toll monitoring
the presence of camera-based systems still raises a lot of
systems can aggregate such data about the type of a vehicle
additional effort in terms of installation, maintenance and also
for providing information about the parking space capacity or
privacy-related problems in real-world scenarios. Furthermore,
for calculating correct toll fees. In this paper, we propose a
the success rate of these systems significantly decreases by
radio-based system which leverages the attenuation of radio
signals for traffic flow detection and vehicle classification Transmitter side
using machine learning algorithms. In contrast to other existing T1 T2 T3
approaches, the proposed system is cost-efficient, easy to in-
stall and does not raise privacy-related issues because it is not
RSSI

based on the evaluation of camera images. Future smart cities

could easily integrate it in delineator posts or traffic lights.
The remainder of this paper is as follows. After discussing
the related work (cf. Sec. II), we present the setup of our
radio-based detection and classification system in Sec. III. In R1 R2 R3 Time
the next section, the system model of our multi-methodical Receiver side
solution approach and its individual components are explained Fig. 1. Example application scenario: structure of the system setup consisting
(cf. Sec. IV). In Sec. V, detailed results for the classification of transmitters and receivers and the effect of passing vehicles on the RSSI
of radio links

978-1-5090-5932-4/17/$31.00 ©2017 IEEE

 is done with the help of different state-of-the-art machine
learning algorithms. A comparison of Logistic Regression
(LR), Neural Networks (NNs) and SVMs for magnetometer-
parameter and accelerometer-based vehicle classification is presented in
Ray tracing
optimization
Field
[14]. The different algorithms achieve similar classification
simulations measurements success rates of about 93%. To further enhance the success
rate of the proposed classification system, we additionally use
realistic ray tracing simulations to evaluate the most suitable
simulation data measurement system settings (e.g., antenna characteristics) and special situa-
special situations data
tions. In general, ray tracing simulations are a popular method
Machine to generate close to reality data, especially in exceptional
Learning situations. The high suitability of ray tracing simulations in
knowledge optimization
the context of vehicular communication is asserted in [15].
The authors give an accuracy comparison of channel-sounder
Vehicle measurements and ray tracing simulations assigning a high
Classification agreement between simulation and measurement data.
Fig. 2. System model of the proposed solution approach using ray tracing III. S ETUP OF THE R ADIO - BASED D ETECTION AND
simulations, measurements in the field and machine learning algorithms.
C LASSIFICATION S YSTEM
The structure of the proposed detection and classification
challenging weather conditions. Among pure camera-based
system is illustrated in Fig. 1. It primarily consists of two
detection systems, there are approaches using laser scanners
wireless key elements: transmitting and receiving units, which
[5], acoustic sensors [6], magnetometers [7] or accelerometers
are positioned on opposite sides of a road spanning a radio
[8]. While the former three approaches suffer from similar
field. It should be pointed out that each transmitter has an
disadvantages like camera-based detection systems, the latter
individual subset of radio links with different receiving nodes.
one requires possibly heavy construction works (pavement cut,
For example, the radio links numbered 1 and 4 are associated
etc). A more convenient approach using a single and portable
to transmitter 1, whereas the links named 6 and 9 are related to
magnetic sensor is presented in [9]. The sensor node is able to
transmitter 3. Typically, the Line of Sight (LOS) propagation
perform a stand-alone detection and classification of vehicles
path for radio transmissions has the largest share of the total
in real time with 99% detection and 60% classification success
transmission power compared to None Line of Sight (NLOS)
rate. In [10] the authors present an approach based on a
paths. Hence, the passage of vehicles through the classifica-
single battery-powered magnetometer classifying vehicles into
tion system results in chronologically fluctuating RSSI traces
three categories. By using Support Vector Machine (SVM)
caused by shadowing. Comprehensive measurements in the
the system reaches a classification accuracy of about 87%.
field have shown that different types of vehicles have specific
Multiple and spatially separated magnetic sensors are used
RSSI fingerprints. In particular, we use this fact to perform
in the classification system proposed in [11]. Due to the
a machine-based vehicle classification for different types of
redundancy of the sensors, the system is able to detect driving
vehicles.
behavior in terms of right-turning or straight-driving vehicles.
Apart from this, the system achieves a classification accuracy
IV. MULTI - METHODICAL APPROACH
of only 83% when using SVM. In order to deal with the
previously mentioned disadvantages and to achieve high classi- For the total system, three different methodological ap-
fication success rates, we propose a detection and classification proaches are brought together: field measurements, ray tracing
system which is integrated in present traffic infrastructure (e.g., simulations and machine learning algorithms (cf. Fig. 2). Data
delineator posts) leveraging the attenuation of radio signals generated in field measurements is used for training and testing
caused by vehicles passing the setup. Essential preparatory purposes as well as for the continuous optimization of the
work with regard to the system setup has been done in a ray tracing simulation parameters. This parameter adjustment
previous project to detect wrong way drivers on motorways is necessary to achieve a high degree of compliance of data
[12], which has a detection rate of over 99% for different traffic generated by simulations with measurement data. In contrast
conditions. Therefore, we adopted the symmetrical setup with to field measurements, ray tracing simulations are used pri-
spatially separated transmitting and receiving units to extend marily to evaluate the most suitable system settings (e.g.,
the system to a high-precision detection and classification antenna characteristics) and secondarily to train and test the
system for parking space accounting on motorway service vehicle classification procedure regarding special situations.
areas based on the analysis of signal attenuation. The fact For example, this includes vehicle types or driving behavior
that vehicles can be considered as obstacles in the signal we have not observed in field measurements yet. In this way,
propagation path is elaborated in [13]. Subsequent to the the classification system can be better prepared on special
evaluation of the signal attenuation, the vehicle classification situations. In the next step, we use the outcome of the vehicle
 TABLE II
classification procedure to recursively optimize our machine- PARAMETERS OF THE FIELD MEASUREMENT SETUP
based classification system. Finally, the overall classification
Parameter Value
process is verified by comprehensive measurements in the Covered area 10 m x 5 m
field. Structure of senders and receivers symmetric
TABLE I Number of senders 3
PARAMETERS OF THE RAY TRACING SIMULATION Number of receivers 3
Types of signal paths direct, diagonal
Parameter Value Number of signals per receiver 2-3
Number of vehicle types 11 Operating frequency 2.4 GHz
Simulation runs per vehicle type 50 RF power 2.5 dBm
Interval steps 0.01 s
Operating frequency 2.4 GHz
Transmitting power 2.5 dBm Tab. I lists relevant parameters used for the simulations. After
Antenna type omni-/directional modeling the simulation scenario, each of its object is provided
Gain of directional antennas 7.1 dBi with appropriate material and motion properties, respectively.
Azimuth 60 deg
Both parameters are of high relevance regarding a realistic
Downtilt 5 deg
Range of transmitters height 0.4 m - 1.2 m
simulation of the signal characteristics at the receiving unit.
Range of receivers height 0.4 m - 1.2 m With reference to the movement behavior, it is possible to
Step size for height 0.2 m distinguish between speed, direction of motion and relative
Distance between transmitters and receivers 7m distance to transmitter and receiver, respectively. Depending
on the simulation scenario, a few motion aspects are limited
A. Ray Tracing Simulations regarding their range of values. For example, it is unlikely that
vehicles drive with very high speeds and in the wrong direction
In order to analyze the effects of different measurement through a parking space. Nevertheless, some of these scenarios
system settings (e.g., characteristics of antennas) and to find are also simulated to evaluate the most suitable settings for
the most suitable parametrization for the live system, we use high classification success rates of the live system in the case
the tools WallMan and ProMan of the ray tracing simulation of special situations.
tool suite WinProp and the Computer-aided Design (CAD)
tool Sketchup for creating and adapting simulation scenarios B. Machine Learning
with highly detailed vehicle models. On the basis of these For the vehicle classification we use different state-of-the-
simulation results, we have evaluated the most appropriate art machine learning algorithms: k-Nearest Neighbor (k-NN)
antenna settings (e.g., antenna type and height for installation). and SVM. These algorithms are trained and tested with the
Fig. 3 illustrates an example simulation scenario based on help of raw data containing chronological RSSI traces from
the proposed system model consisting of multiple transmitters field measurements and simulations. With this approach, the
and receivers, which is passed by a SUV. Here, the colored knowledge for classification is consecutively enhanced with
box represents the location-dependent received power of the a large set of real-world data as well as a small portion of
signal transmitted by transmitter 2. Apparently, the passage of simulation data. Beside the RSSI traces length information is
vehicles has significant effects on the received signal strength. used as second feature for classification. The training of the
classification system has been done supervised with the two
Transmitter 1 Receiver 1 labels passenger car and truck. To control the performance of
the classification process a standard five-fold cross-validation
was used.
C. Measurements in the Field
5m

For validating the classification success rate of the pro-

Transmitter 2
Receiver 2 posed machine learning-based system, we use new data sets
generated by field measurements. In order to achieve a high
variety of vehicle types as well as a low influence of the
environment on the classification, we have chosen a driveway
for a parking space next to the motorway A9 in Germany as
location for our field measurement setup. The most important
Transmitter 3 Receiver 3 parameters of this setup are summarized in Tab. II. Basically,
the setup consists of a total of six delineator posts equipped
RSSI with directional antennas, microcontroller boards for signal
[dBm]
- 68 - 64 - 60 - 56 - 52 processing and a power supply. While one half of those
delineators serves as transmitters, the remaining delineators
Fig. 3. Example ray tracing scenario illustrating the shadowing effects caused are used as receivers. In addition to the signal reception,
by a SUV passing the proposed classification system.
 -40 50
RSSI [dBm]
Significant difference concerning the impact
-60 of ground-reflected radio waves on the RSSI
-80
Considerably longer level 40
drops for semi trucks with trailers
-100
0 100 200 300 400 500 600

RSSI [dBm]
-40 30
RSSI [dBm]

8 dBm
-60

-80 Significantly shorter and lower 20

level drops for passenger cars
-100
0 100 200 300 400 500 600
Time [ms] 10

Fig. 4. Example field measurement data of three trucks and three passenger
cars. 0
Pass. car Pass. car Truck Truck
the receivers are also responsible for signal processing and (sim.) (meas.) (meas.) (sim.)
forwarding of the processed data to the master gateway. After Fig. 5. Analysis of the influence of ground-reflected radio waves on the RSSI
collecting all data, the master gateway performs the main part for passenger cars and trucks based on measurement and simulation data.
of the whole classification procedure.
higher RSSI level. To check the suitability of RSSI traces as
V. R ESULTS OF THE P ERFORMANCE E VALUATION
a feature for the classification process, we have analyzed the
In this section, we present the results achieved with the impact of ground-reflected radio waves across the street on the
proposed classification system. First, we examine the signif- received signal strength. Fig. 5 shows the magnitudes of RSSI
icance of ground-reflections on the received signal strength level drops due to the shadowing caused by passenger cars
with the help of measurement and simulation data. Then, and trucks for measurement and simulation data, respectively.
we present the results using the features length information Apparently, the mean drop of signal strengths caused by
and RSSI traces for classification. Finally, we consider the passenger cars is about 8 dBm higher compared to the one
classification success rate of the proposed system as our main caused by trucks (cf. measurement data in Fig. 5). This results
key performance indicator. from the significantly different shapes and distances of the car
bodies of various vehicle types and the surface of a street.
A. Impact of Ground-reflected Radio Waves
As a consequence, the RSSI traces of radio links across the
Subsequently, the results of the proposed classification sys- street differ for various vehicle types due to varying impacts
tem for use of field measurement data are presented. Fig. 4 of ground-reflections caused by the street. This assumption is
shows example RSSI traces caused by the passage of trucks confirmed with the help of simulation results also shown in
and passenger cars. Obviously, the developing of the RSSI Fig. 5. Finally, these results reveal the high suitability of RSSI
traces differs in terms of duration and magnitude of signal traces as feature for vehicle classification.
dropout depending on the vehicle type passing the measure-
ment system. In particular, trucks cause a significantly higher B. Classification Results using Field Measurement Data
and considerably longer drop of the RSSI levels compared to In order to achieve a high classification accuracy for various
passenger cars. In addition, the temporal developing indicates types of vehicles, the machine learning algorithms have been
whether a truck with or without a trailer is passing the trained and tested via five-fold cross validation. By using RSSI
classification system. For example, one trace in the upper part traces as exclusive feature, the cross validation leads to an
of Fig. 4 illustrates the passage of a semi truck with a trailer. accuracy of 98.68% ± 0% for k-NN and 98.68% ± 0.31%
The reason for the temporary peak of signal strength at the for SVM (cf. Tab. IV). If length information is used as an
time of about 300 ms is the free space between the semi truck additional feature, an accuracy of 99.56% ± 0% for k-NN
and the trailer. At this point, there is almost a LOS signal path and 99.47% ± 0.20% for SVM is achieved. Tab. III contains
between transmitter and receiver resulting in a significantly the results of the proposed classification system for the two

TABLE III
C LASSIFICATION SUCCESS RATE OF K -NN AND SVM REGARDING THE LABELS PASSENGER CAR AND TRUCK WITH FIELD MEASUREMENT DATA

Length RSSI Traces Length & RSSI Traces

Label Vehicle type Test samples Rec. rate Test samples k-NN SVM Test samples k-NN SVM
Passenger car 541 100.0% 60 100.0% 100.0% 60 100.0 % 100.0%
Passenger Small van 12 100.0% 12 100.0% 91.67% 12 100.0% 100.0%
car Van 100 92.00% 20 100.0% 100.0% 20 100.0% 100.0%
Transporter 107 85.98% 27 96.30% 96.30% 27 96.30% 92.59%
Bus 6 100.0% 6 100.0% 100.0% 6 100.0% 100.0%
Truck
Truck 503 94.04% 103 98.06% 99.03% 103 100.0% 100.0%
Overall success rate 1269 95.82% 228 98.68% 98.68% 228 99.56% 99.12%
labels passenger car and truck when using only aggregated the help of sensor fusion and the evaluation of further signal
length information, the RSSI traces as exclusive feature or the characteristics even more.
combination of both features. The classification results have
ACKNOWLEDGMENT
been evaluated with MATLAB and validated with the help of
Part of the work on this paper has been supported by the German Federal
RapidMiner [16]. Ministry for Economic Affairs and Energy as part of the cooperation project
TABLE IV between Wilhelm Schröder GmbH, TU Dortmund and FH Dortmund (grant
C ROSS VALIDATION : C LASSIFICATION SUCCESS RATES OF K -NN AND agreement number ZF4038101DB5), and by Deutsche Forschungsgemein-
SVM REGARDING THE RECOGNITION OF THE LABELS PASSENGER CAR schaft (DFG) within the Collaborative Research Center SFB 876 “Providing
AND TRUCK USING RSSI TRACES AS EXCLUSIVE FEATURE . Information by Resource-Constrained Analysis”, project B4 “Analysis and
Communication for Dynamic Traffic Prognosis”.
Training set Test set k-NN SVM
R EFERENCES
S2, S3, S4, S5 S1 98.68% 98.68%
S1, S3, S4, S5 S2 98.68% 98.68% [1] A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi, “Internet
of things for smart cities,” IEEE Internet of Things Journal, vol. 1, no. 1,
S1, S2, S4, S5 S3 98.68% 98.25%
pp. 22–32, Feb 2014.
S1, S2, S3, S5 S4 98.68% 98.68% [2] L. A. Klein, M. Mills, and D. Gibson, “Traffic detector handbook,”
S1, S2, S3, S4 S5 98.68% 99.12% Federal Highway Administration, Tech. Rep., 2006. [Online].
Available: http://www.fhwa.dot.gov/publications/research/operations/its/
06108/06108.pdf
Due to the symmetrical structure of the proposed system, a [3] J. Zhang, F. Y. Wang, K. Wang, W. H. Lin, X. Xu, and C. Chen, “Data-
highly accurate determination of vehicle length information is driven intelligent transportation systems: A survey,” IEEE Transactions
possible, which can also be used for vehicle classification. on Intelligent Transportation Systems, vol. 12, no. 4, pp. 1624–1639,
Dec 2011.
However, for a few vehicle types (e.g., transporters) the [4] J.-W. Hsieh, S.-H. Yu, Y.-S. Chen, and W.-F. Hu, “Automatic traffic
sole evaluation of length information results in comparably surveillance system for vehicle tracking and classification,” IEEE Trans-
low classification success rates. The reason for this is the actions on Intelligent Transportation Systems, vol. 7, no. 2, pp. 175–187,
June 2006.
wide variety of vehicles associated with only two labels. [5] X. Zewei, W. Jie, and C. Xianqiao, “Vehicle recognition and classifica-
For example, there are transporters differentiating in terms tion method based on laser scanning point cloud data,” in Transportation
of shape and size. Therefore, the members of a vehicle Information and Safety (ICTIS), 2015 International Conference on, June
2015, pp. 44–49.
type may differ significantly from each other. In comparison [6] J. George, L. Mary, and R. K. S, “Vehicle detection and classification
to the sole evaluation of length information (95.82% total from acoustic signal using ANN and KNN,” in Control Communication
accuracy), overall classification success rates of 98.68% can and Computing (ICCC), 2013 International Conference on, Dec 2013,
pp. 436–439.
be achieved by using RSSI traces. In order to further improve [7] S. Taghvaeeyan and R. Rajamani, “Portable roadside sensors for vehicle
the classification success rate, the combination of both features counting, classification, and speed measurement,” IEEE Transactions on
is used for classification (cf. Tab. III). Apparently, the overall Intelligent Transportation Systems, vol. 15, no. 1, pp. 73–83, Feb 2014.
[8] W. Ma, D. Xing, A. McKee, R. Bajwa, C. Flores, B. Fuller, and
classification success rates increase from 98.68% to 99.56% P. Varaiya, “A wireless accelerometer-based automatic vehicle classifica-
(k-NN) and from 98.68% to 99.12% (SVM). In particular, tion prototype system,” IEEE Transactions on Intelligent Transportation
the advantage of this approach is revealed with regard to the Systems, vol. 15, no. 1, pp. 104–111, Feb 2014.
[9] S. Cheung, S. Coleri, B. Dundar, S. Ganesh, C.-W. Tan, and P. Varaiya,
recognition of trucks. If RSSI traces are used exclusively, the “Traffic measurement and vehicle classification with single magnetic
classification system achieves an accuracy of 98.06% (k-NN) senso,” Transportation Research Record: Journal of the Transportation
and 99.03% (SVM). Instead, the usage of both features leads Research Board, vol. 1917, pp. 173–181, 2005. [Online]. Available:
http://dx.doi.org/10.3141/1917-19
to classification success rates of 100% when using k-NN and [10] H. Tafish, W. Balid, and H. H. Refai, “Cost effective vehicle classi-
SVM, respectively. Finally, the overall classification success fication using a single wireless magnetometer,” in 2016 International
rates are slightly improved by combining both features. Wireless Communications and Mobile Computing Conference (IWCMC),
Sept 2016, pp. 194–199.
[11] S. Taghvaeeyan and R. Rajamani, “Portable roadside sensors for vehicle
VI. C ONCLUSION counting, classification, and speed measurement,” IEEE Transactions on
Intelligent Transportation Systems, vol. 15, no. 1, pp. 73–83, Feb 2014.
In this paper, we presented a radio-based approach for traffic [12] S. Haendeler, A. Lewandowski, and C. Wietfeld, “Passive detection
flow detection and vehicle classification which combines ray of wrong way drivers on motorways based on low power wireless
tracing simulations, machine learning and measurements in communications,” in 2014 IEEE 79th Vehicular Technology Conference
(VTC Spring), May 2014, pp. 1–5.
the field. First, the impact of ground-reflected radio waves on [13] M. Boban, T. T. V. Vinhoza, M. Ferreira, J. Barros, and O. K. Tonguz,
the RSSI traces were examined revealing their high suitability “Impact of vehicles as obstacles in vehicular ad hoc networks,” IEEE
as feature for our classification system. The results proved Journal on Selected Areas in Communications, vol. 29, no. 1, pp. 15–28,
January 2011.
the high performance of the proposed classification system [14] D. Kleyko, R. Hostettler, W. Birk, and E. Osipov, “Comparison of
with regard to the classification of two labels. By combining machine learning techniques for vehicle classification using road side
the two features RSSI traces and vehicle length information sensors,” in 2015 IEEE 18th International Conference on Intelligent
Transportation Systems, Sept 2015, pp. 572–577.
the classification system achieves an accuracy of more than [15] J. Nuckelt, T. Abbas, F. Tufvesson, C. Mecklenbrauker, L. Bernado,
99%. Moreover, the proposed classification system has low and T. Kurner, “Comparison of ray tracing and channel-sounder mea-
deployment and maintenance costs recommending it for ob- surements for vehicular communications,” in Vehicular Technology Con-
ference (VTC Spring), 2013 IEEE 77th, June 2013, pp. 1–5.
taining relevant data in a smart city context. In future work, [16] RapidMiner GmbH. [Online]. Available: https://rapidminer.com/
we will increase the accuracy of the proposed system with