A Low Cost Modular Radio Tomography System for Bicycle and Vehicle Detection and Classification
AA Low Cost Modular Radio Tomography Systemfor Bicycle and Vehicle Detection and Classification
Marcus Haferkamp, Benjamin Sliwa and Christian Wietfeld
Communication Networks Institute, TU Dortmund University, 44227 Dortmund, Germanye-mail: { Marcus.Haferkamp, Benjamin.Sliwa, Christian.Wietfeld } @tu-dortmund.de Abstract —The advancing deployment of ubiquitous Internetof Things (IoT)-powered vehicle detection and classificationsystems will successively turn the existing road infrastructure intoa highly dynamical and interconnected Cyber-physical System(CPS). Though many different sensor systems have been proposedin recent years, these solutions can only meet a subset ofrequirements, including cost-efficiency, robustness, accuracy, andprivacy preservation. This paper provides a modular systemapproach that exploits radio tomography in terms of attenuationpatterns and highly accurate channel information for reliableand robust detection and classification of different road users.Hereto, we use Wireless Local Area Network (WLAN) andUltra-Wideband (UWB) transceiver modules providing eitherChannel State Information (CSI) or Channel Impulse Response(CIR) data. Since the proposed system utilizes off-the-shelf andpower-efficient embedded systems, it allows for a cost-efficientad-hoc deployment in existing road infrastructures. We haveevaluated the proposed system’s performance for cyclists andother motorized vehicles with an experimental live deployment.In this concern, the primary focus has been on the accuratedetection of cyclists on a bicycle path. However, we also haveconducted preliminary evaluation tests measuring different mo-torized vehicles using a similar system configuration as for thecyclists. In summary, the system achieves up to 100% accuracyfor detecting cyclists and more than 98% classifying cyclists andcars.
Accepted for presentation in: 2021 Annual IEEE International Systems Conference (SysCon)2021 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses,including reprinting/republishing this material for advertising or promotional purposes, collecting new collected worksfor resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
I. I
NTRODUCTION
Comprehensive and reliable Intelligent Transportation Sys-tems (ITSs) are a crucial feature for emerging smart citiesas the continuing increase in road traffic will noticeablyexhaust the capacity of existing traffic systems [1]. In manycases, constructional measures for expanding a traffic system’scapacity are not an option, so traffic flow optimization is theonly valuable solution resulting in data-driven ITSs. By con-tinuously gathering specific information for different vehicletypes, those systems enable more comprehensive traffic flowoptimization than approaches providing only coarse indicatorslike traffic flow and traffic density. Hence, those systems mustmeet several conditions at once, including a high detectionand classification accuracy in real-time, even for challengingweather conditions. Moreover, they should provide energy-efficient, low-maintenance, and thus cost-efficient operationwhile being privacy-preserving. The compliance with thosedemands is highly relevant, particularly for mass deploymentsused in smart city applications. However, most of the existingsolutions lack at least one of these criteria, disqualifying themfor large-scale deployments. Hence, we present a modular and highly integrated Wire-less Sensor Network (WSN) installation for vehicle detectionand classification that leverages both attenuation and high-dimensional channel information. The central assumption isthat each vehicle induces type-specific radio channel patterns( fingerprints ), allowing for accurate vehicle detection andclassification. Hereto, we use different state-of-the-art MachineLearning (ML) models suitable for deployment to off-the-shelf Microcontroller Units (MCUs) for implementing a highlyautomated classification process. Thus, our system fulfills thepreviously mentioned requirements for smart city applications,i. e., high detection and classification accuracy, robustnessagainst challenging weather conditions, cost-efficiency, andprivacy-preservation. The initial Wireless Detection and Warn-ing System (WDWS) has exploited the attenuation of radiolinks induced by passing vehicles to detect wrong-way driverson motorways [2]. Subsequently, this approach has beensuccessively adopted for a fine-grained and ML-based vehicleclassification of multiple vehicle classes [3].This paper proposes a modular and highly integrated radio-based detection system, allowing for cost-efficient mass de-ployments in urban road infrastructure. As an example, Fig. 1illustrates the proposed system’s use for automated detectionand classification of cyclists and vehicles in an urban scenario.The contribution of this paper is as follows: • Presentation of a low-cost, power-efficient, and modularradio tomography system for vehicle detection and clas-sification exploiting highly accurate channel information • Performance comparison of state-of-the-art machine
RSSI t tRSSI
Fig. 1. Example application: Using the modular radio tomography systemto detect and classify heterogeneous road users in an urban setting. a r X i v : . [ ee ss . SP ] F e b WB RX PowerWLAN CSIWLAN RSSI R e l a t i v e A m p li t ude s Time
Fig. 2. Example raw and smoothed radio fingerprints of a cyclist for WLANRSSI, WLAN CSI, and UWB received signal power. For WLAN CSI, eachline indicates a subcarrier’s relative amplitude. learning methods—Artificial Neural Network (ANN),Random Forest (RF), Support Vector Machine (SVM)—for two classification tasks • In-depth suitability analysis of parameters extracted from
WLAN CSI and UWB CIR channel information
After giving an overview of related work in Sec. II, we providethe modular and radio-based classification system approachin Sec. III, the methodology in Sec. IV, and present theperformance analysis in Sec. V.II. R
ELATED W ORK
In this section, we provide a brief overview of various sen-sor technologies used for vehicle detection and classificationsystems. Hereafter, we focus on related radio-based sensorsystems and corresponding ML models.
A. Sensor Technologies for Detection and Classification
Each vehicle detection and classification system can beclassified either as intrusive or non-intrusive . While theformer system type represents the original system design andimplies expensive roadwork for installation and maintenance(e. g., pavement cut), the latter is typically well-suited forlarge-scale deployments due to less extensive effort.Specifically, systems categorized as intrusive are: Weigh inMotion (WIM) [4], induction loops [5], [6], fiber Bragg gratingsensors [7], vibration sensors [8], and piezoelectric sensors [9].Contrary, there is a variety of non-intrusive sensor technologiesused for detection and classification systems, which includesacoustic sensors [10], [11], inertial sensors [12], [13], vision-based [14], [15] as well as radio-based systems. In the follow-ing, we discuss radio-based approaches in more detail. B. Radio-based Sensor Systems
Radio-based approaches leverage radio tomography andRadio Tomographic Imaging (RTI) [16] for conducting de-tection and classification tasks. Such systems are WSNsranging from simple one-link setups to collaborative multi-technology systems exploiting different radio technologies.The basic assumption of radio tomography is that objectsof different shapes and materials lead to characteristic radiosignal patterns. The resulting radio fingerprint can be usedfor several kinds of object detection and tracking by takingsnapshots over time (cf. Fig. 2).The Received Signal Strength Indicator (RSSI) is a granularmeasure representing an estimate of the total received signalstrength provided by most transceiver modules. For instance,the RSSI is used in WSNs equipped with Bluetooth LowEnergy beacons for vehicle detection and classification [17],achieving a detection and classification accuracy of up to 98%and 97% for three vehicle types, respectively. In [3], the au-thors propose an RSSI-based multi-link vehicle classificationsystem capable of conducting binary classifications with morethan 99% and more fine-grained seven-type classifications withmore than 93% accuracy assessing the RSSI of each radio link.In contrast to RSSI, WLAN CSI provides frequency-specificdetails regarding a radio channel. In general, OrthogonalFrequency-Division Multiplexing (OFDM)-based radio sys-tems estimate CSI for compensating a radio link’s interfer-ences to reconstruct the original symbols. In particular, theCSI describes the estimated impact of the channel on bothamplitude and phase of each subcarrier in the Long-TrainingField (LTF) of a received packet. The total size of the CSIdepends on the number of transmit antennas, receive antennas,and subcarriers, whereas the latter varies between 64 and 512subcarriers depending on the used channel bandwidth.The great potential of CSI becomes apparent when lookingat various research activities. For instance, Adib et al. applylocalization and tracking of moving objects behind a wall orclosed doors. Furthermore, this approach also allows for de-tecting simple gestures performed behind a wall [18]. Keenanet al. utilize this potential to distinguish three forms of humanfalling enabling privacy-preserving monitoring by healthcareapplications. The proposed system achieves a balanced ac-curacy of 91%, determining intended fall-like activities likesitting down and harmful ones such as walking-falls [19].Although UWB is primarily used for indoor and outdoorlocalization, Sharma et al. compare the feasibility of WLANCSI and UWB for device-free Human Activity Recognition(HAR) [20]. According to the presented results, UWB outper-forms WLAN CSI using an ML-based classification for threedifferent activities.Concerning traffic monitoring, Won et al. leverage CSIusing two laptops equipped with WLAN Network InterfaceControllers (NICs), detecting and classifying a total of fivedifferent vehicle types. The proposed classification systemtransforms the low-pass filtered and Principal ComponentAnalysis (PCA)-treated CSI data into image data, which serves ata Acquisition
Road
S1R1 S2R2
Schematics Field DeploymentSetup
Data Preprocessing
SmoothingNormalizationFeature Extraction
Measurement Point S C A m p li t ude Data Analysis
Single-type DetectionMulti-type Classification
Data Exploitation
On-Site Exploitation
Parking Space AccountingTraffic MonitoringPredictive Road WorkSmart Parking P $ Toll Collection P CO Emission Control
Global Exploitation
Fig. 3. Overall system architecture model for a low-cost and modular radio tomography system for road user detection and classification. Radio fingerprintsare gathered, preprocessed, evaluated using ML algorithms, and exploited for different ITS applications. as input for a Convolutional Neural Network (CNN), leadingto average vehicle detection and classification accuracies of99.4% and 91.1%, respectively [21].Instead of utilizing only a single radio technology, Wanget al. propose a Collaborative Sensing Mechanism (CSM)-based real-time vehicle detection and classification systemcombining power-efficient magnetic sensors and power-hungrycameras. While the low-cost magnetic sensors are runningcontinuously for vehicle detection, the latter is usually in low-power mode and awake only for real-time vehicle classificationand counting. This collaborative WSN approach reaches aclassification accuracy of at least 84% for the vehicle typesbicycle (98.84%), car (95.71%), and minibus (84.38%)[22].Usually, CSI is processed within the transceiver modulesand, therefore, not directly accessible in most off-the-shelfdevices. Hence, recent research has originated tools for extract-ing CSI from specific WLAN NICs [23], [24]. However, using
Espressif ESP32
MCUs in our modular radio tomogra-phy system, we can directly access CSI through the officialfirmware Application Programming Interface (API) [25].
C. Machine Learning
In recent years, the availability of numerous differentlyscaling ML algorithms has promoted their use in many ap-plication areas, including the cognitive optimization of radio-based applications. For vehicle detection and classification,the focus is on supervised learning models such as ANN,RF [26], and SVM [27]. In contrast, more modern and complexML approaches—such as Deep Neural Networks (DNNs)—areused less frequently due to their demand for large datasets.Moreover, ML models perform differently, mainly dependingon the number of considered vehicle classes, the systemdeployment’s environment, and the used WSN, differentiatingin the number of links, sensor technologies, etc.III. S
OLUTION A PPROACH
In this section, we explain the proposed solution approachand its components. For a better overview, Fig. 3 illustrates theoverall system architecture model containing four basic pro-cessing steps: data acquisition in the live system deployment, data preprocessing—including smoothing, normalization, andfeature extraction—, ML-based data analysis considering spe-cific classification tasks, and data exploitation as required byvarious ITS applications.
Data Acquisition:
Due to its data-driven nature, real-world traces of road users—e. g., bicycles and motorizedvehicles—are gathered using a low-cost and modular radio-based WSN setup. We evaluate two radio communicationtechnologies: WLAN CSI and UWB (cf. Fig 4). We uti-lize
Espressif ESP32
MCUs to access WLAN CSI andcustom-made Printed Circuit Boards (PCBs), combining a
Decawave DWM1000
UWB transceiver module and an
ARMCortex M3
MCU [28]. Both MCUs provide the channeldata via Universal Serial Bus (USB) interface for furtherprocessing.
Data Preprocessing:
The raw WLAN CSI and UWB CIRdata passes a three-step process cascade, including smoothing,normalization, and feature extraction. We conduct the datasmoothing with a one-dimensional Gaussian filter evaluatingdifferent values for the Gaussian kernel’s standard deviation σ . Hereafter, the smoothed data is normalized such thatthe values are bound to the range [0 , ( min-max-scaling ).While we perform the smoothing to minimize the impact ofscattered outliers—e. g., due to fading in the radio channel—the normalization enables high compatibility with the used MLalgorithms ( feature scaling ). The last step is the extraction ofmultiple descriptive statistical features. In total, we have de- UWB Transceiver ARM MCUESP32 MCU
Fig. 4. WLAN CSI (left) and UWB (right) transceiver modules evaluatedin the low-cost and modular radio tomography system. ived more than 20 attributes for the ML-based classification.
Data Analysis:
In the third process step, we feed thepreprocessed data as input for two data analysis options.While one option targets the detection of only one specificvehicle type, the other one is required to detect and classifymultiple vehicle types correctly. For instance, we performedthe coarse-grained detection task along a cycle path countingcyclists. The latter application is more relevant for urbanenvironments revealing heterogeneous road users, includingpedestrians, cyclists, and several motorized vehicles.
Data Exploitation:
Finally, one could use the obtaineddata analysis results to provide multiple ITS-related serviceseither within a specific site (on-site exploitation) or on a largescale (global exploitation). Possible applications for on-siteexploitation are parking space accounting, traffic monitoring,or toll collection. In contrast, analysis data acquired frommultiple sensor deployments within a region can serve as inputfor smart parking, emission control, and predictive road work.IV. M
ETHODOLOGY
This section provides details regarding the modular radiotomography system’s parameters, the vehicle taxonomies as-sumed for the classification task, and in-depth informationabout the ML models we have applied in the evaluation step.
A. Field Deployment Setup
Tab. I summarizes the essential system parameters of theproposed radio-based detection and classification system. Wehave comparably installed WLAN CSI and UWB transceivermodules in the field deployment setup. Nevertheless, somedifferences face the transmission power or the antenna char-acteristics induced either by the transceiver modules’ design orthe radio technology. Moreover, there is a variation concerningthe distances between transmitter and receiving nodes formeasuring cyclists and motorized vehicles. We have gatheredradio fingerprints along a cycle path and a busy one-lane road,respectively. Since most captured fingerprints are related tocyclists (995 traces), this paper’s primary focus is on detectingthese—which can be interpreted as a binary classificationof bicycle and non-bicycle . For this reason, we also havecaptured idle traces, i. e., there is a Line of Sight (LOS)
TABLE IM
AIN P ARAMETERS OF THE P ROPOSED M ODULAR R ADIO T OMOGRAPHY S YSTEM U SING
WLAN CSI
AND
UWB
FOR B ICYCLE D ETECTION AND V EHICLE C LASSIFICATION . Parameter Radio Technology
WLAN CSI UWB
Transmission power 20 dBm 9.3 dBmOperating frequency 2.4 GHz 6.5 GHzSampling frequency 80 Hz 40 HzAntenna type Directional OmnidirectionalAntenna gain 5-7 dBi —Antenna height 1m 1mNumber of radio links 1 1Distance TX ↔ RX (cycle path) 4m 4mDistance TX ↔ RX (road) 7m 7m
Non-Car-likeCar-like ( ) Bicycle (473)
Van ( FHWA 3 ) SUV ( FHWA 2 ) Passenger car ( FHWA 2 ) Fig. 5. Taxonomies: Vehicle classes and sample numbers used in the multi-type classification task. We considered balanced subsets for car-like, bicycle,and idle samples. between transmitter and receiver. Hereafter, we also evaluatethe proposed system’s applicability for a more fine-graineddetection and classification task of three types: idle, cyclist,and car-like vehicles (cf. Fig. 5).
B. Machine Learning
For the detection and classification, we utilize multiple mod-els that have different implications for the achievable accuracyand resource efficiency. These considered models are chosenwith respect to the findings of [3], which yielded that often lesscomplex classification models achieve better accuracy resultsthan cutting edge methods that would require a significantlyhigher amount of training data for achieving a comparableperformance level due to the curse of dimensionality . • Artificial Neural Networks (ANNs) [29] aim to mimiccore functions of the human nervous system and havereceived tremendous attention within various scientificcommunities in the context of deep learning . Thesemodels can be implemented as a sequence of matrixmultiplications with element-wise node activations. Theresulting memory size of ANNs is determined by theircorresponding network architecture. Due to the usage offloating-point arithmetic, ANNs are less popular for beingused on highly resource-constrained IoT platforms suchas ultra low power microcontrollers. • Random Forests (RFs) [26] are ensemble methods thatbase their decision making on the joint consideration ofa number of random trees. Each tree is trained on arandom subset of the features and a random subset ofthe training data. The layer-wise descent within the treesis based on binary decision making, whereas the valueof a single feature is compared to a learned threshold.Due to condition-based decision making, RFs can beimplemented in a highly resource-efficient manner as asequence of if/else statements. Varying the numberof trees and the maximum tree depth allows to controlthe memory usage of RFs. • Support Vector Machines (SVMs) [27] learn a hyper-plane for separating data points in a multidimensionalspace through minimization of a specific objective func-tion. The hyperplanes are chosen for each feature thatmost members of one of two classes are on each of thehyperplane sides. We apply the one-vs-all strategy forusing SVM for multi-class learning problems.n order to assess the generalizability of the achievedclassification results, we apply a k = 10 -fold cross-validationand investigate the variance of the model performance. Hereby,the overall data set D is divided into k subsets {D , ..., D k } .In each iteration i , D i is chosen as the training set D train forthe model, and the remaining subsets jointly compose the testset D test .All data analysis tasks are carried out using the high-levelLIghtweight Machine Learning for IoT Systems (LIMITS)framework [30] for automating Waikato Environment forKnowledge Analysis (WEKA) [31] evaluations. In addition, itallows exporting C/C++ code of trained prediction models.V. P
ERFORMANCE A NALYSIS
In this section, we discuss the results for bicycle detectionand multi-type vehicle classification using the proposed modu-lar radio tomography system. Essentially, we show the resultsfor both the WLAN CSI and the UWB radio modules.
A. Bicycle Detection
As mentioned in Sec. IV-A, this paper’s primary focusis on accurately detecting cyclists on a cycle path, i. e.,differentiating bicycles and non-bicycles (idle). Nonetheless,we also provide results for a more fine-grained classificationtask in the following section. Tab. II shows the classificationresults for the separately analyzed channel parameters acquiredfor WLAN CSI and UWB using the ML models ANN, RF, andSVM. Concerning WLAN CSI,
RSSI is the dominant channelparameter leading to the best classification results—for allscores. A possible explanation is that the WLAN transceivermodule evaluates multiple channel parameters for calculating asingle and significant indicator. Similarly, one channel parame-ter is most relevant when using the UWB transceiver modules:the quotient of the estimated
First Path Power (FPP) and the
Channel Impulse Response (CIR) power , where the latter isthe sum of the magnitudes’ squares from the estimated highestpower portion of the channel. Using this extracted parameter
FPP/CIR and ANN, we achieve a bicycle detection (binaryclassification) accuracy of 100%.
TABLE IIB
ICYCLE D ETECTION : R
ESULTS FOR
WLAN CSI
AND
UWB
USING
ANN, RF,
AND
SVM
WITH A
FOLD C ROSS V ALIDATION (CV)
Model Score WLAN CSI UWB
Value [%] Param. Value [%] Param.ANN Accuracy 99.27 ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± f : Filter size, FC : Ratio of first path signal power and CIR power, R : RSSI UWB C l a ss i fic a t i on A cc u r a cy [ % ] WLAN CSI R S S I H _ A M P S C H _ A M P S C S _ A M P S C F P P / C I R R
X P F P P F P P - C I R C I RH _ A M P Fig. 6. Bicycle detection: Five most relevant channel parameters for WLANCSI and UWB, respectively. We evaluated each parameter separately usingRF and 10-fold CV.
CIR : CIR power,
FPP : First path signal power,
H AMP :Amplitudes of HT-LTF subcarriers,
RSSI : Received signal strength indicator,
RXP : Estimated received signal power,
S AMP : Amplitudes of STBC-HT-LTFsubcarriers, SC : Subcarrier. Fig. 6 illustrates the five most relevant channel parametersof WLAN CSI and UWB for bicycle detection using RF. Aspreviously discussed, the
RSSI (WLAN CSI) and the quotient
FPP/CIR (UWB) are the most significant channel parametersfor correctly detecting cyclists. While the investigated UWBparameters lead to small deviations regarding the classificationaccuracy, there is at least 2% lower accuracy comparing RSSIand the remaining WLAN CSI parameters.Fig. 7 depicts the ten most significant extracted statisticalfeatures for
RSSI and
FPP/CIR . For both systems, we canidentify small differences in their relative feature importancedistributions. Again for WLAN CSI, there is a single dominantfeature ( kstat ), whereas we cannot determine such a superiorone regarding UWB. R e l a t i v e F ea t u r e I m po r t an c e Statistical Features
UWBWLAN CSI
Fig. 7. Bicycle detection: Feature importance for the ten most relevantextracted statistical features for WLAN CSI and UWB using RF. iqr :interquartile range, kstat : k-static, mad : median absolute deviation, q05 : th quantile, q95 : th quantile, sem : standard error of mean, std : standarddeviation, tmean : trimmed mean, tvar : trimmed variance, var : variance. ubcarrier Groups C l a ss i f i c a t i on A cc u r a cy [ % ] G1 G2 G6G3 G4 G7 G8ANNRFSVM
ANNRFSVM
WLAN CSI
Fig. 8. Bicycle detection: Classification accuracy using different subcarrieramplitudes as input for ANN, RF, SVM, respectively. For a better overview,we have grouped adjacent subcarriers.
Finally, Fig. 8 presents the significance of different WLANCSI subcarrier amplitudes for the given binary classificationtask utilizing ANN, RF, and SVM. For a better overview,we have split adjacent SCs into eight groups. We can statea frequency-specific relevance of these SCs regarding theclassification accuracy. In particular, the SCs of G1 (SCs1-8) are less suitable than those of the remaining groups.Furthermore, we can observe comparably high accuraciesusing ANN and RF, but consistently lower ones using SVM. B. Multi-Type Vehicle Classification
This section provides an outlook on the modular radiosystem’s applicability for multi-type vehicle classification. Fora total of three evaluated categories—idle, bicycle (non-car-like), and car-like—Tab. III lists the classification results forWLAN CSI and UWB using ANN, RF, and SVM, respec-tively. Contrary to the cyclist detection task, there are at leasttwo predominant channel parameters for each system.Concerning WLAN CSI, the Legacy Long Training Field(LLTF) subcarriers’ amplitudes ( L ) are most suitable usingANN; instead, the STBC-HT-LTF subcarriers’ amplitudes ( S )are more crucial when applying RF. There are two relevant pa-rameters when using SVM: the LLTF subcarriers’ amplitudes( L ) and the amplitudes of the nd subcarrier in the HT-LTFtraining field ( H SC ).Focusing on the classification results achieved for UWB,there are also two major channel parameters: the amplitudesof all raw CIR accumulator data ( A ) and the amplitudes ofaccumulator sample 15 ( A ). When comparing the classifi-cation results for both systems, we can state a considerableperformance gap for the benefit of WLAN CSI. We notethat we have gathered traces of car-like vehicles on a busyone-lane road, implying a more substantial distance betweensending and receiving nodes than for measuring cyclists, which TABLE IIIM
ULTI - TYPE V EHICLE C LASSIFICATION : R
ESULTS FOR
WLAN CSI
AND
UWB
USING
ANN, RF,
AND
SVM
WITH A
FOLD CV Model Score WLAN CSI UWB
Value [%] Param. Value [%] Param.ANN Accuracy 98.23 ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± SC52 (f0) 91.17 ± (f0)Precision 97.86 ± SC52 (f0) 92.13 ± (f0)Recall 97.46 ± ± (f0)F-Score 97.39 ± SC52 (f0) 91.29 ± (f0) A : Amplitudes of all CIR accumulator samples, A : Amplitudes of CIR accumulatorsample 15, f : Filter size, H SC : HT-LTF SC 52 amplitudes, L : LLTF SCsamplitudes, S : STBC-HT-LTF SCs amplitudes may significantly affect the UWB transceiver modules usingomnidirectional antennas.Fig. 9 illustrates the relevance of different channel pa-rameters gathered from WLAN CSI and UWB regarding athree-type classification using RF. Concerning the results,several WLAN CSI channel parameters lead to classificationaccuracies in the range of 97% to 98%. Contrary, the overallclassification performance is notably worse, using any ofthe evaluated UWB parameters. The most suitable UWBparameter U AMP results in about 4% lower accuracy thanthe fifth most relevant WLAN CSI parameter
H AM P SC .Furthermore, we achieve considerably different accuracy levelsreaching from about 93% down to 87% using the five mostrelevant UWB parameters. We assume that the divergentantenna types and sampling rates of the used WLAN CSIand UWB transceiver modules (cf. Tab. I) may cause thisperformance gap. C l a ss i f i c a t i on A cc u r a cy [ % ] S _ A M P H _ A M P U _ A M P AS U _ A M P AS U _ A M P AS U _ A M P F P P H _ A M P M EA N H _ A M P S C H _ A M P S C UWBWLAN CSI
Fig. 9. Multi-type vehicle classification: Five most relevant channel param-eters for WLAN CSI and UWB, respectively. We evaluated each parameterseparately using RF and 10-fold CV. AS : Accumulator sample index, FPP :First path signal power,
H AMP : Amplitudes of HT-LTF SCs,
S AMP :Amplitudes of STBC-HT-LTF SCs, SC : Subcarrier, U AMP : Amplitudes ofCIR accumulator samples
I. C
ONCLUSION
In this paper, we presented a novel bicycle detection andmulti-type vehicle classification system that exploits highlyaccurate channel parameters provided by WLAN CSI andUWB. Compared to existing traffic detection and classificationsystems, the proposed modular radio tomography system isprivacy-preserving, robust against challenging weather condi-tions, and cost-efficient. Using real-world data from extensivefield measurements, we have analyzed its applicability fortwo classification tasks with different state-of-the-art machinelearning models. Regarding the detection of cyclists, whichwe conducted as a binary classification task, an accuracy ofmore than 99% can be achieved for both radio technologiesWLAN CSI and UWB, using ANN, RF, and SVM, respec-tively. Furthermore, we have evaluated the proposed system’sperformance for a multi-type classification gaining more than98% accuracy.In future work, we will improve the system’s accuracy bycorrelating multiple radio links and extracting different radiochannel parameters. Moreover, we will obtain additional sam-ples of various vehicles involving challenging urban settings—e. g., in a downtown area with groupings of vehicles—anddifferent weather conditions to strengthen the overall systemperformance. In the long term, the full detection and classi-fication process, including the process steps discussed in thispaper, should run self-sufficiently on the utilized MCUs.A
CKNOWLEDGMENTThis work has been supported by the PuLS project (03EMF0203B) fundedby the German Federal Ministry of Transport and Digital Infrastructure(BMVI) and the German Research Foundation (DFG) within the CollaborativeResearch Center SFB 876 “Providing Information by Resource-ConstrainedAnalysis”, projects A4 and B4. We would like to thank Tugay Onat for hishelpful support conducting the field measurements. R EFERENCES[1] B. Sliwa, T. Liebig, T. Vranken, M. Schreckenberg, and C. Wietfeld,“System-of-systems modeling, analysis and optimization of hybrid ve-hicular traffic,” in , Orlando, Florida, USA, Apr 2019.[2] S. Haendeler, A. Lewandowski, and C. Wietfeld, “Passive detectionof wrong way drivers on motorways based on low power wirelesscommunications,” in , 2014, pp. 1–5.[3] B. Sliwa, N. Piatkowski, and C. Wietfeld, “The channel as a traffic sen-sor: Vehicle detection and classification based on radio fingerprinting,”
IEEE Internet of Things Journal , Mar 2020.[4] L. A. Klein, M. K. Mills, and D. P. Gibson, “Traffic detector handbook:Third edition - volume I,” 2006.[5] L. Wu and B. Coifman, “Improved vehicle classification from dual-loop detectors in congested traffic,”
Transportation Research Part C:Emerging Technologies , vol. 46, pp. 222 – 234, 2014.[6] J. Lamas, P.-M. Castro-Castro, A. Dapena, and F. Vazquez-Araujo,“Vehicle classification using the discrete fourier transform with trafficinductive sensors,”
Sensors , vol. 15, 10 2015.[7] M. Al-Tarawneh, Y. Huang, P. Lu, and D. Tolliver, “Vehicle classificationsystem using in-pavement fiber bragg grating sensors,”
IEEE SensorsJournal , vol. 18, no. 7, pp. 2807–2815, 2018.[8] Z. Ye, H. Xiong, and L. Wang, “Collecting comprehensive trafficinformation using pavement vibration monitoring data,”
Computer-AidedCivil and Infrastructure Engineering , vol. 35, no. 2, pp. 134–149, 2020.[9] S. Rajab, M. O. Al Kalaa, and H. Refai, “Classification and speed esti-mation of vehicles via tire detection using single-element piezoelectricsensor,”
Journal of Advanced Transportation , vol. 50, no. 7, pp. 1366–1385, 2016. [10] J. George, L. Mary, and R. S., “Vehicle detection and classification fromacoustic signal using ANN and KNN,” 12 2013.[11] C. Daniel and L. Mary, “Fusion of audio visual cues for vehicleclassification,” in , 2016, pp. 1–4.[12] C. Xu, Y. Wang, X. Bao, and F. Li, “Vehicle classification using animbalanced dataset based on a single magnetic sensor,”
Sensors , vol. 18,p. 1690, 05 2018.[13] W. Ma, D. Xing, A. McKee, R. Bajwa, C. Flores, B. Fuller, andP. Varaiya, “A wireless accelerometer-based automatic vehicle classi-fication prototype system,”
Intelligent Transportation Systems, IEEETransactions on , vol. 15, pp. 104–111, 02 2014.[14] A. J. Siddiqui, A. Mammeri, and A. Boukerche, “Towards efficient ve-hicle classification in intelligent transportation systems,” in
Proceedingsof the 5th ACM Symposium on Development and Analysis of IntelligentVehicular Networks and Applications , ser. DIVANet ’15. New York,NY, USA: Association for Computing Machinery, 2015, p. 19–25.[15] K. Liu and G. Mattyus, “Fast multiclass vehicle detection on aerialimages,”
IEEE Geoscience and Remote Sensing Letters , vol. 12, no. 9,pp. 1938–1942, 2015.[16] C. R. Anderson, R. K. Martin, T. O. Walker, and R. W. Thomas, “Radiotomography for roadside surveillance,”
IEEE Journal of Selected Topicsin Signal Processing , vol. 8, no. 1, pp. 66–79, Feb 2014.[17] M. Bernas, B. Płaczek, and W. Korski, “Wireless network with bluetoothlow energy beacons for vehicle detection and classification,” in
Com-puter Networks , P. Gaj, M. Sawicki, G. Suchacka, and A. Kwiecie´n,Eds. Cham: Springer International Publishing, 2018, pp. 429–444.[18] F. Adib and D. Katabi, “See through walls with WiFi!”
SIGCOMMComput. Commun. Rev. , vol. 43, no. 4, pp. 75–86, Aug. 2013.[19] R. M. Keenan and L. N. Tran, “Fall detection using Wi-Fi signalsand threshold-based activity segmentation,” in , 2020, pp. 1–6.[20] S. Sharma, H. Mohammadmoradi, M. Heydariaan, and O. Gnawali,“Device-free activity recognition using ultra-wideband radios,” in , 2019, pp. 1029–1033.[21] M. Won, S. Sahu, and K. Park, “Deepwitraffic: Low cost wifi-basedtraffic monitoring system using deep learning,” in ,2019, pp. 476–484.[22] R. Wang, L. Zhang, K. Xiao, R. Sun, and L. Cui, “EasiSee: Real-timevehicle classification and counting via low-cost collaborative sensing,”
IEEE Transactions on Intelligent Transportation Systems , vol. 15, no. 1,pp. 414–424, Feb 2014.[23] D. Halperin, W. Hu, A. Sheth, and D. Wetherall, “Tool release: Gathering802.11n traces with channel state information,”
SIGCOMM Comput.Commun. Rev. , vol. 41, no. 1, p. 53, Jan. 2011.[24] Y. Xie, Z. Li, and M. Li, “Precise power delay profiling with commoditywi-fi,”
IEEE Transactions on Mobile Computing , vol. 18, no. 6, pp.1342–1355, 2019.[25] M. Atif, S. Muralidharan, H. Ko, and B. Yoo, “Wi-ESP—A tool for CSI-based device-free Wi-Fi sensing (DFWS),”
Journal of ComputationalDesign and Engineering , 05 2020.[26] L. Breiman, “Random forests,”
Mach. Learn. , vol. 45, no. 1, pp. 5–32,Oct. 2001.[27] C. Cortes and V. Vapnik, “Support-vector networks,”
Mach. Learn. ,vol. 20, no. 3, pp. 273–297, Sep. 1995.[28] J. Tiemann and C. Wietfeld, “Scalability, real-time capabilities, andenergy efficiency in ultra-wideband localization,”
IEEE Transactions onIndustrial Informatics , vol. 15, no. 12, pp. 6313–6321, 2019.[29] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,”
Nature , vol. 521,no. 7553, pp. 436–444, 5 2015.[30] B. Sliwa, N. Piatkowski, and C. Wietfeld, “LIMITS: Lightweight ma-chine learning for IoT systems with resource limitations,” in , Dublin, Ireland,Jun 2020, Best paper award.[31] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, andI. H. Witten, “The WEKA data mining software: An update,”