density-based statistical clustering: enabling sidefire...

Research ArticleDensity-Based Statistical Clustering Enabling SidefireUltrasonic Traffic Sensing in Smart Cities

Volker Luumlcken Nils Voss Julien Schreier Thomas Baag Michael GehringMatthias Raschen Christian Lanius Rainer Leupers and Gerd Ascheid

Institute for Communication Technologies and Embedded Systems (ICE) RWTH Aachen University Aachen Germany

Correspondence should be addressed to Volker Lucken lueckenicerwth-aachende

Received 18 August 2017 Accepted 6 December 2017 Published 4 January 2018

Academic Editor Mehmet Yildirimoglu

Copyright copy 2018 Volker Lucken et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Traffic routing is a central challenge in the context of urban areas with a direct impact on personal mobility traffic congestionand air pollution In the last decade the possibilities for traffic flow control have improved together with the correspondingmanagement systems However the lack of real-time traffic flow information with a city-wide coverage is a major limiting factor foran optimum operation Smart City concepts seek to tackle these challenges in the future by combining sensing communicationsdistributed information and actuation This paper presents an integrated approach that combines smart street lamps with trafficsensing technology More specifically infrastructure-based ultrasonic sensors which are deployed together with a street lightsystem are used for multilane traffic participant detection and classification Application of these sensors in time-varying reflectiveenvironments posed an unresolved problem for many ultrasonic sensing solutions in the past and therefore widely limited thedissemination of this technology We present a solution using an algorithmic approach that combines statistical standardizationwith clustering techniques from the field of unsupervised learning By using a multilevel communication concept centralized anddecentralized traffic information fusion is possible The evaluation is based on results from automotive test track measurementsand several European real-world installations

1 Introduction

In the context of the Internet of Things (IoT) the integraltransformation process of urban infrastructure into intel-ligent and connected devices plays an important role forfuture mobility Cities experience an increasingly high levelof road congestion due to the focused traffic coming alongwith progressing urbanization This congestion induces ahigh social economical and environmental cost From aEuropean Union (EU) perspective it is amounted to aboutone percent of the GDP [1] and predicted to increase byabout 50 percent by the year 2050 [2] From an environmentalpoint of view carbon dioxide emissions and air pollutionespecially locally in cities are a significant social and healthaspect [3] With this new unprecedented level of trafficdensity offline and online traffic planning including liverouting techniques becomes a requisite for larger cities beinguseful both for end users and for city planners Especially inthe case of real-time traffic monitoring and the subsequent

information provision to drivers and public routing systemsproper conditioning of this traffic information in terms ofaccuracy area coverage and density is paramount Trafficcongestion can only be effectively reduced if this informationbase which is used for the routing decisions is deliveredwith a sufficient level of quality in these dimensions Thenwith a city-wide traffic status large-scale end-to-end routingoptimizations can be performedThe results can be deployedto the vehicle driver with provision methods such as in-car information systems or centralized traffic guidance fortriggering the individualrsquos routing reaction If this provisionis not performed properly either due to an insufficienttraffic information quality or improper consideration ofgroup effects in the optimization even detrimental effectscan occur [4] In this article these requirements on trafficinformation quality and density are in focus building abase for future routing decisions Many currently used trafficsensing techniques for creating this live data basis will onlyhave a negligible positive impact on the routing decisions

HindawiJournal of Advanced TransportationVolume 2018 Article ID 9317291 15 pageshttpsdoiorg10115520189317291

2 Journal of Advanced Transportation

if they are distributed sparsely and do not achieve sufficientcoverage

In general two main types of techniques achieving morecomprehensive traffic information have been establishedin the last years A first category is the fixed or mobileinfrastructural distribution of sensors like inductive loopsand traffic camerasThe second type is represented by systemsrelying on end-user based distributed information for exam-ple movement data from smartphone users or informationgathered using V2X communication approaches Details andcharacteristics of several techniques are discussed in thispaper considering not only metrics like traffic informationquality and performance in different scenarios but also pos-sible privacy issues which may arise with certain techniques

The novel ultrasound-based traffic sensing techniquewhich is the focus of this paper belongs to the category ofinfrastructural sensing A significant difference to existingultrasonic sensor systems is the placement of the sensors ina so-called sidefire setup This means that the sensors aremounted in a height of at least threemeters and face the streetsideways with downtilted sensors while being positionedon the side of the street Previously existing systems weremainly single-lane ultrasonic sensors measuring the heightprofile top-down for example from a special sensor gantryabove the street with a simple first-reflection processingThe monitoring of multiple lanes is possible with our setupwhile at the same time the requirements in terms of algorith-mic evaluation and processing are increased in comparisonto these previously existing systems However with theseimprovements our approach allows operation in a highlyreflective acoustic environment which is present in urbanareas and permits simply mounting the sensor on the sideof the street This is enabled by a new approach with acombination of statistical signal processing clustering andinference algorithms for traffic participant object detection

This sidefire ultrasonic traffic sensing technique is partof a development which focuses on intelligent infrastructuresolutions more specifically on intelligent street lights Anexemplary integrated setup is shown in Figure 1 Theselight modules can be used to replace inner components ofstreet lamps for example by retrofitting the widespread typeof high-pressure sodium-vapor lamps Then they can actas a flexible platform for many applications by employingthe processing communication and light control actuationcapabilitiesWith the transition to LED-based lighting whichis triggered by high energy costs and several EU directives[5] a large proportion of the existing street lamps will haveto be replaced in the next decade offering potential for anintegration of Smart City solutions The seamless integrationinto infrastructure allows an almost complete coverage ofrelevant urban traffic to provide the required coverage densityof traffic information Furthermore the major problem ofpower supply for sensing devices is alleviated as the lamps areconnected to the mains power which eliminates the need forenergy-constrained and expensive battery-powered devices

The rest of this article is organized as follows Section 2provides the related work of urban traffic sensing techniquesin general and ultrasonic sensing in specific Section 3describes the system architecture together with the specific

Figure 1 Integrated street lamp head with a single ultrasonic sensorhead

scenario and typical sensor setup Section 4 provides thealgorithmic approach to the sensor signal processing systemand object detection inference In Section 5 the evaluationmethodology and framework for reference data are pre-sented Section 6 gives results of real-world measurementsand provides an evaluation from the perspective of detectionand classification together with a state-of-the-art technologycomparison Section 7 then concludes the article

2 Related Work

Ultrasonic traffic participant detection falls into the field oftraffic sensing with a large number of techniques availablein research and in practical application In this sectionwe therefore provide a state-of-the-art overview from twoperspectives First we give a comprehensive overview oftraffic monitoring techniques in general with a discussionof their performance strengths and weaknesses in order toshow the technological gaps and potential In the second partspecific technical approaches in ultrasonic traffic sensingwhich are rare both in concepts and in practical implemen-tations are given This application-centric perspective showsthe drawbacks and unresolved problems in current state ofthe art

21 Traffic Monitoring Techniques from a Broad View In theintroductory section two main categories of traffic informa-tion acquisition systems were introduced These are on theone hand infrastructure-based sensing systems working ina temporarily or permanently fixed location On the otherhand the type of end-user assisted sensing systems incor-porates distributed information acquired from the driverstheir technical devices and the vehiclersquos internal sensorsitself In the following a state-of-the-art overview togetherwith a technological comparison is given for nowadays mostrelevant traffic monitoring techniques of both categoriesA condensed overview of the properties is also given inTable 1

Themostwidely used approach to traffic sensing is the useof infrastructure-based solutions which represents a mainlycentralized approach In general authority of the data iskept with the operator and thus several of the problemscoming with user-centric techniques do not arise Howeverthe sensing systems have to be deployed over the whole cityor at least at critical locations in terms of traffic Achieving therequired density for an extensive urban traffic monitoring istherefore paramount which brings along high effort and costas a centralized solution

Journal of Advanced Transportation 3

Table 1 Overview of state-of-the-art traffic monitoring techniques

Sensing technology Characteristics (oplus⊖) ReferencesInfrastructural sensing techniques

Infrared sensors(passive and active)

oplus Vehicle class and speed information acquisition is possible

[10 11]oplusMultilane operation possible (active sensors)⊖ Susceptible to bad weather conditions⊖ Regular cleaning required (active sensors)

Camera-based systemsoplusHigh detection and classification accuracy

[6 7]⊖ Privacy and authorization issues⊖ Susceptible to bad weather and light conditions

In-pavement sensors(inductive piezoelectric and strain)

oplusHigh detection accuracy with axle counting [8 9]⊖ Invasive and costly installation due to in-pavement integration

Directional high-end radar sensorsoplusMultilane multiobject traffic participant detection

[10 12 13]oplus Directional and velocity information⊖ Cost and complexity of solutions

Top-down ultrasonic and radar sensorsoplus Low cost and simple mode of operation

[10 14ndash18]⊖ Only single-lane operation⊖Mounting above lane required for example on sensor gantry

Proposed sidefire ultrasonic sensingoplusMultilane operation possibleoplus Sidefire mounting allows seamless integration into environment⊖ No directional information available in sidefire configuration

User-based distributed traffic monitoringtechniques

Crowdsourcing systems(smartphones and navigation devices)

oplusHigh number of devices[21ndash23]⊖ Privacy concerns

⊖ Susceptibility to attacks and data manipulation

V2X-based cooperative systems oplus Combination with in-vehicle information and decisions [24]⊖ Low support and prevalence

Camera-Based Systems Frequently camera-based systems areused for traffic monitoring Here we focus on systems withan automated analysis that is based on Computer Visionalgorithms With camera-based systems a high detectionaccuracy plus a vehicle classification can be achieved Dur-ing nighttime further solutions are required for exampleartificial illumination of the road and infrared or thermalimaging Only in bad visibility conditions as fog snow orsmog the performance is limited [6] The main concernwith these systems however is the issue of privacy arisingwith video cameras in the public domain On the one handthe personal privacy of the citizens is affected on the otherhand governmental regulations might even forbid the use ofcameras in public areas especially in this high-coverage caseAlso even with only a sparse coverage of video cameras inthe public space acceptance levels amongst citizens are low[7]

In-Pavement Sensors Sensors in the urban area embeddedinto the pavement asphalt are often used in city intersectionareas for traffic light control The main in-pavement sensortechnologies are inductive loops piezoelectric sensors andstrain gauges [8] They can also enable high-accuracy trafficmeasurements together with a classification by counting

the axles of vehicles [9] As a downside the installation ofpavement-integrated sensors is highly invasive and costlyand requires per-lane road work with a possible temporarydisturbance of city traffic

Infrared Sensors Passive and active infrared sensors arenonintrusive sensor types which can be mounted on theside or directly above the street They offer detection withadditional speed and vehicle class information in somerealizations With active sensors multilane operation ispossible However infrared sensors are susceptible to badweather conditions like rain snow or fog Furthermoreactive infrared sensors require regular cleaning which mightlead to lane closure during maintenance [10 11]

Radar Systems From the basic object detection principleradar sensors represent the technology most comparable toultrasonic sensors in the field of traffic sensing High pulserepetition rates and ranges together with the possibility ofobtaining directional information together with velocity dataallow the monitoring of multiple traffic lanes and objectsindependent of the weather conditions [10 12 13] As amajor downside the significant cost of the radar systemincluding front-end and signal processing can compromise


a city-wide system deployment for a sufficient coverage withthese sensors

Top-Down Ultrasonic and Radar In the field of radar andultrasonic sensors several top-down sensor solutions areavailable which require the placement above a specific streetlane for single-lane car profile measurements [10 14ndash18]These techniques allow obtaining the complete height profileof vehicles for detection and classification but require agantry or bridge over the street for mounting the sensorswhich is infeasible in many situations Moreover only asingle lane can be monitored with the existing solutions[10] For the ultrasonic sensors the processing is stronglysimplified as with the available systems only the distanceto the first relevant reflection and not the complete impulseresponse is evaluated The significantly lower propagationspeed and resulting decay time of the impulse responsestogether with the requirements for a sufficient pulse repeti-tion rate (typically ge10Hz) of ultrasonic sensors can lead toself-interference and also to issues with fast-moving trafficparticipants [18] For ultrasonic sensors previous studieshave shown that the impact of weather effects such as windsnow and rain is only marginal [19 20]

Crowdsourcing Systems For the second category of end-userbased distributed information systems for trafficmonitoringone of the most present techniques is the crowdsourcing oftraffic monitoring to end-user smartphones and navigationdevices in order to obtain real-time and high-density trafficflow information [21 22] This approach is used in severalsystems and solutions for example by ldquoGoogle Mapsrdquo withthe product ldquoGoogle Trafficrdquo or by ldquoTomTomHD Trafficrdquo Asa central advantage no investments in additional sensors arerequired and the very high prevalence of these applicationson end-user devices can be exploited by actively trackingthe devices However the live tracking of the users is highlycritical in terms of privacy aspects Furthermore thesesystems are susceptible to manipulative techniques and fakeinformation due to trust issues with the user base which hasbeen demonstrated before [23]

V2X Communications With the emerging field of V2X(vehicle-to-everything) communications which is currentlyfocused on on V2V (vehicle-to-vehicle) and V2I (vehicle-to-infrastructure) techniques new possibilities for trafficinformation acquisition and exchange arise Togetherwith in-vehicle information for a direct delivery of inferred routingdecisions to the driver and in a fusion with available travelplans on the car navigation device an optimized informa-tion exchange and routing is enabled Nevertheless thesetypes of systems are still to be realized in mass markettogether with suitable application scenarios and it will taketime until they are sufficiently established amongst everydayvehicles Moreover reservations with regard to privacy andtrust are also an important aspect for these systems Inliterature it is shown that with a sufficient prevalence ofcooperative V2V techniques in cars together with a highrelaying communication range dense traffic informationcan be obtained [24] Therefore vehicular communication

techniques are a promising approach for the future Certainlyas long as the density and proportion of vehicles supportingV2X communications is too low a diverse traffic monitoringapproach which in parallel comprises infrastructural sensingtechniques has to be pursued for several decades

22 Specific Advances in Ultrasonic Traffic Detection Inthe previous introductory Section 21 the basic top-downultrasonic sensing technology has already been introducedHowever the practical use is limited as additional mountingefforts are required if no infrastructure like bridges abovethe street or traffic light poles at intersections are availableMoreover only single-lane detection is possible in this regardTherefore these solutions are limited in their practical appli-cability

There also exist initial concepts that use ultrasonic sensorsmounted on a lateral street position with a so-called sidefireorientation facing the opposite street side either horizontallyor in a downtilted position [19 25 26] In the case of hor-izontal tilt even a first simple multilane detection has beenrealizedHowever all these concepts performonly an analysisof the first-reflection distance instead of the whole impulseresponse which highly limits the practical use in a varietyof scenarios The urban scenario is a highly reflective time-varying environment where first reflections of other objectssuch as moving trees parking cars and the scattering of theroad surface are always present in the distance range withthe relevant traffic participants for detection the distanceregion-of-interest for measured signal reflections of objectsIn addition the horizontally facing type of sidefire sensorscould be obstructed by other objects and traffic participantsin urban scenarios due to their low mounting height

3 System Concept and Architecture

Sensor systems have specific requirements regarding thesetup scenario and working environment they are beingoperated in In this chapter the system setup and applicationconcept together with the generic requirements for ourproposed ultrasonic traffic sensing system are given Thescenario stays the same as initially proposed in Section 1 aSmart City street lamp system platform Nevertheless a widevariety of installation and application possibilities are givenalso without any ties to this integration concept

31 Sensor System Setup An exemplary setup of the smartstreet lamp platform was already shown with Figure 1 It fea-tures a single mounted ultrasonic sensor attached to the lamphead or the pole and an embedded system integrated into acustomized LED-based lamp head The digital core systemcontains modules for realizing processing and communica-tion capabilities enabling distributed processing in a local-level group of lamps in a street by building a dynamic meshnetwork and also allowing communication on the globaland cloud level by employing mobile network machine-to-machine type communication (M2M) functionalities Thecentral component for the signal processing capabilitiesconsists of a Texas Instruments Sitara system-on-chip (SoC)including an ARM Cortex-A8 core Typical examples of


Θ

Θ Downtilt angle Beam width

Figure 2 Downtilted sidefire sensor configuration in typical urbanenvironment with two-lane street (with downtilt angle Θ and beamwidth 120572)

possible applications with this system are location-basedcontent delivery attachment of modules for environmentalor traffic sensing applications and energy efficiency opti-mizations of street lighting Due to the combination ofinfrastructure with an extensible platform the systems canbe widely distributed amongst the city as almost all centralareas are covered by street lamps Furthermore many of theconstraints on processing and communications which aretypically imposed for Smart City traffic monitoring systemsin IoT context can be relaxed for example data and refreshrate restrictions due to battery power [5]

The typical seamless integration of the ultrasonic sensingconcept into an urban street area is shown in Figure 2In the so-called downtilted sidefire setup the sensor ismounted in a height between 3 and 8m which is above thetypical height of vehicles in urban areas The sensor headis then oriented diagonally downwards the street with adowntilt angle Θ between 30∘ and 80∘ The ultrasonic sensortransducer together with the casing has a cone-shaped beamcharacteristic with an opening angle 120572 typically between20∘ and 80∘ depending on the scenario and lane coveragerequirements In the exemplary scenario shown in Figure 2themultitude of reflective elements in the urban environmentalready becomes apparent For example the first reflectionsand scattering of the street surface can arrive earlier thanreflections from relevant objects on distant lanes to thesensor Additionally reflective objects can exhibit short- orlong-term time variation like parked cars or trees moving inthe wind In these cases simple distance-measurement basedsensor techniques would already fail

32 Ultrasonic Sensor Platform Module In order to providethe capabilities for ultrasonic sensing the core system isextended with a custom sensor platform module which isshown in Figures 3 and 4 It provides four ultrasonic sensorchannels for real-time arbitrary waveform generation andrecording in half-duplex mode The reception quality of theultrasonic impulse response is with an input stage SNR of

above 105 dB in the ultrasonic band The analog drivingcircuit for the transducer is integrated into the ultrasonictransducer head itself and is adapted to the capabilities andcharacteristics of the transducer

Control of the attached components signal generationand preprocessing capabilities is realized with an integratedXilinx Artix-7 series FPGA It provides communicationcapabilities to the core processing system over an Ethernetlink which is also used for the power supply of the sen-sor system extension Together with the core system GPSincluding the high-precision PPS signal is used for a timingsynchronization and exact coordination of multiple sensorsin the vicinity of a street Interference can therefore becoordinated by a global optimization of the sensor timingpatterns if the sensor heads are transmitting in the samefrequency band

33 General Signal Structure For simplicity the signal struc-ture in the single-sensor case is discussedThe transmit signal119904119879(119905) consists of repeated sequences of the base pulse ℎ119894(119905)with an overall duration 119879rep yielding a pulse repetition rateof 119891rep = 119879minus1rep This is also shown in Figure 5 119879rep is typicallychosen between 50 and 150ms to ensure low self-interferencelevels between the different impulse response blocks Therepetition time should in general be larger than the RT60reflections decay time of the acoustic channel [27 p 98] TheRT60 time is a common measure in indoor acoustics butcan also be applied in outdoor scenarios where it typicallyyields much shorter values The general pulse sequence isthen defined as follows

119904119879 (119905) =infinsum119899=0

ℎ119899 (119905 minus 119899 sdot 119879rep) (1)

The base pulses ℎ119894(119905) itself consist of two specific duplexerphases At the beginning the transmit duplex mode is usedand a transmit pulse part of length 119879wnd is emitted Thenthe duplex mode is switched to receive mode and records thechannel impulse response for the rest of the period119879repminus119879wndwithout further transmission For the investigations in thisarticle the transmit pulse is chosen to be a windowed signalwith a carrier frequency of 119891119888 = 40 kHz and an envelopewindow ℎwnd119894(119905) with length 119879wnd = 2ms as a tradeoffbetween time and frequency resolution for the narrowbandsignal The 119894-th base pulse is then given by the following

ℎ119894 (119905) = ℎwnd119894 (119905) sdot Re 1198901198952120587119891119862119905 if 0 le 119905 le 119879wnd0 otherwise (2)

With the use of identical repeated transmit pulses inour case ℎ119894(119905) = ℎ(119905) forall119894 isin N A Blackman window [28]function is chosen for ℎwnd119894(119905) because it does not exhibitdiscontinuities at the beginning and end of the time domainwindow while exhibiting a high stop-band attenuation [29]

4 Processing Concept and Algorithms

The ultrasonic traffic sensing technique presented in thispaper is based on a combination of statistical analysis


Ultrasonic sensor processing platformFPGA Xilinx ARTIX-7 XC7A100T Cat 5 sensor cable

(PowerTxRxControl)

Ultrasonictransducer

Ethernet + PoE

Connection to mainstreet lightsystem-on-chip

Sensor head withdriverduplex circuit

ADC

DACWaveformgenerator

andacquisition

Signalpreprocessing

Timing synchronization

Controland

managementprocessor

DDR3 memory

PoE power supply

Industrial Ethernet

PHY


ADC

DAC

Figure 3 General structure of sensor platform including the FPGA and attached ultrasonic sensors with driving circuits

Figure 4 Custom sensor platform which is integrated as anextension module into the intelligent street light

0

Tran

smit

signa

l am

plitu

de

Transmitreceive signal with modulated carrierTransmit envelope with windowed pulses(Blackman window)Received impulse response envelope

Pulse repetition time TＬＪ

Transmitpulse length

TＱＨ＞

1 20Normalized time (1TＬＪ s)

Figure 5 Generalized signal structure with transmitreceiveduplexing mode corresponding envelopes and modulated signals

and clustering techniques for object candidate detectionTogether with a special signal chain structure objects repre-senting traffic participants can be detected together with anextraction of further characteristics and object properties Inthis section the model for the received signal is given firstThen the characteristics of the reflective environment aremodeled and empirically analyzed Based on these findingsstatistical hypothesis testing is performed during operationand the resulting statistical information is combined with amodified clustering algorithm for object boundary detection

Then the system concept for the extraction of further char-acteristics is given and the reduced-complexity embeddedimplementation for real-time operation is discussed

41 Received Signal Model and Preprocessing With the gen-eral signal structure described in Section 33 repeated mea-surements of the acoustic channel are performed with theultrasonic sensor system Based on the impulse responses asthe one-dimensional received signals which consist of a largenumber of reflections in the acoustic environment objectdetection and further analyses can be performed togetherwith the time-of-flight distance information

The received signal 119904(119905) from the sensor head is recordedcontinuously and then sliced into blocks of length 119879repyielding a series of impulse responses 119904119894(119905119863) with 119894 beingthe 119894th measurement interval of length 119879rep The parameter119905119863 is the time position of the specific impulse responsewhich is mapped to an equivalent time-of-flight distance119889 = 119905119863 sdot 119888 given the speed of sound 119888 = 343ms inthe air medium at normal conditions Due to the round-trip the real distance to the reflective object is half theequivalent time-of-flight distance By acquiring informationon the current air temperature via sensor measurements orweather information 119888 can be compensated for the currentconditions During the initial period of 119904119894(119905) with 0 le 119905119863 le119879wnd which is the transmission duplexmode no useful signalis acquired while the region 119879wnd le 119905119863 le 119879rep represents themain signal of interest for further analysisThe single impulseresponse slices are then given by the following

119904119894 (119905119863) = 119904 (119894 sdot 119879rep + 119905119863) (3)

The equivalent time-discrete representation is

119909119894 (119899119863) = 119904 (119894 sdot 119873rep119879 + 119899119863 sdot 119879) (4)

with the sampling time 119879 and the discrete-time repetitioninterval length119873rep = 119879rep119879

As a first step of preprocessing of the sliced receivesignal on the FPGA each impulse response is convolvedwith a bandpass filter ℎBP(119905) centered at the transmit carrierfrequency 119891119862 with a bandwidth of 6 kHz for allowing theanalysis of Doppler-shifted signals with typical urban vehiclevelocities Then a time-discrete Hilbert transformH [30 31]


123456789

10

Dist

ance

dim

ensio

n

10

20

30

40

50

60

Dist

ance

dim

ensio

n

0 5 2010 15Time dimension (s) (with TＬＪ = 100 ms)

minus80 minus60 minus10minus70 0minus50 minus20minus40 minus30

Normalized envelope function amplitude (dB)

impu

lse d

elay

tim

e (m

s)

equi

vale

nt d

istan

ce (m

)

Figure 6 Two-dimensional envelope function mapping of receivedsignal 119890119877119894(119899119863) (amplitude normalized and in logarithmic scale)

is performed in order to acquire the complex-valued analyticsignal

119909119894 (119899119863) = (119904119894 lowast ℎBP) (119899119863) + 119895 sdotH (119904119894 lowast ℎBP) (119899119863) (5)

together with the instantaneous amplitude or real-valuedenvelope signal

11989010158401015840119894 (119899119863) = 1003816100381610038161003816119909119894 (119899119863)1003816100381610038161003816 (6)

This envelope signal 11989010158401015840119894 (119899119863) is further processed tofacilitate object detection in the following Only in caseof Doppler-based motion and velocity estimation time-frequency information such as the instantaneous phase andfrequency has to be incorporated Both its two discretedimensions 119894 being the index of the impulse response (fromnow on called time dimension with an equivalent samplingrate of 119891rep) and 119899119863 being the distance-equivalent impulseresponse time (from now on called distance dimension withan equivalent sampling rate 119891119878 = 1119879) are derived froma single original time dimension in the received signal byslicingTherefore the distance dimension is fixed and limitedto 0 le 119899119863 le 119873rep minus 1 while the causal time dimension inreal-world application is extending with 119891rep With this two-dimensional mapping object detection algorithms becomeapplicable which consider a vicinity in both dimensionsThismeans that effects of an object creating a peak at severalneighboring distance and time points are now properlyrepresented as neighbors in the 2D space in contrast to theoriginal one-dimensional recorded sensor signal

An example for the envelope signal 11989010158401015840119877119894(119899119863) is given inFigure 6 with a normalized logarithmic color mapping ofthe amplitude for improved visualization The equivalentdistance and impulse response time mappings are shown onthe ordinate for reference purposes In this plot the scatteringand reflections from the street surface can already be seenin a distance range of 4 to 6m in a comb-like pattern Alsoaround the time positions at 7 s 14 s 15 s 17 s and 19 s strongreflections from objects passing the sensor are visible

42 Signal Statistics and Standardization Given the highcomplexity of the reflective patterns in the sensor data the

statistics of the signal need to be described properly in orderto perform an outlier or anomaly detection which wouldindicate the presence of objects in the sensor field The goalis to incorporate all noise components (measurement noiseand other acoustic emissions in the ultrasonic band) and alsothe typical static and time-varying reflections (eg movingleaves of a tree caused by wind and reflection patterns ofthe street) for every specific distance point of the signalinto the statistics This allows the classification of segmentsof newly acquired impulse responses in terms of the datafollowing the a priori distribution called base distributionwhich was estimated in the past or whether an outlier ispresent

In contrast to typical binary decision problems here onlythis base distributionwithout any presence of an object can beestimated representing the null hypothesis The alternativehypothesis of object presence is however different for everyobject This problem can be solved using parametric ornonparametric hypothesis testing techniques for properlyrepresenting the underlying base distributions of the signalpoints and testing the outlier probabilities However for thereal-time implementation on the embedded system of thesensor setup we present a simpler and less computationallycomplex approach that is based on the distance-wise stan-dardization of the signal based on the calculated statistics ina predefined time window in the past These standardizationtechniques are often used in the field of data science asa preprocessing step for feature scaling The distance-wisestandardized envelope signal 11989010158401015840119894 (119899119863) is calculated based onthe information from the statistics memory with a windowlength of 119873119879119908 time dimension steps in the past (equivalentto a time range of 119873119879119908 sdot 119879rep in continuous time) yielding1198901015840119894 (119899119863)

1198901015840119894 (119899119863) = 11989010158401015840119894 (119899119863) minus 120583119899119863 (119894)120590119899119863 (119894) (7)

= 11989010158401015840119894 (119899119863) minusWindowed mean 120583119899119863 (119894) for distance 119899119863⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞119873minus1119879119908sum119894119896=119894minus119873119879119908 11989010158401015840119896 (119899119863)

radic(119873119879119908 minus 1)minus1sum119894119898=119894minus119873119879119908 [11989010158401015840119898 (119899119863) minus 119873minus1119879119908sum119894119896=119894minus119873119879119908 11989010158401015840119896 (119899119863)]2

⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟Windowed standard deviation 120590119899119863 (119894) for distance 119899119863

0 le 119899119863 le 119873rep minus 1 119894 ge 119873119879119908

(8)

Under the simplified assumption that the underlyingprocess of 11989010158401015840119894 (119899119863) (with no objects present) for a givendistance 119899119863 is an iid Gaussian process 119883119899119863 sim N(120583119899119863 1205902119899119863)the specific distance ranges of 1198901015840119894 (119899119863) will follow the standardnormal distribution 119884119899119863 sim N(0 1) The assumption isespecially found to be valid for direct reflections in the closerrange shown in studies from other acoustic environmentslike underwater [32] As a general approximation it will laterbe used for approximatively calculating the initial detectionthresholds for a given false alarm rate

For the object detection only the right tail of the distri-bution is of interest for outlier detection as shadowing effectsare not considered Therefore the complete standardizeddistribution data 1198901015840119894 (119899119863) is clipped for values below zero


resulting in the final envelope 119890119894(119899119863)with statistical standard-ization

119890119894 (119899119863) = max (1198901015840119894 (119899119863) 0) (9)

This yields a variable with a standard normal distributionleft-censored at 0 The resulting probability distributionfunction (PDF) 119891(119909) can be written as follows

119891 (119909) =

1radic2120587119890minus11990922 + 1

2120575 (119909) 119909 ge 00 otherwise (10)

and the cumulative distribution function (CDF) 119865(119909) is

119865 (119909) =

12erf (

119909radic2) + 1

2 119909 ge 00 otherwise

(11)

where erf(119909) is the Gauss error function The first terms of119891(119909) and119865(119909) are similar to a scaled half-normal distribution[33] with the addition of the clipped values being statisticallycensored and not truncated

To summarize 119890119894(119899119863) is now used for all further objectdetection and processing together with the clustering tech-niques in the following Simply put it yields the positiveoutlier level of a newly acquired data point by the distanceto the calculated mean in a measure of the multiple ofstandard deviations This is based on the statistical historyin a specific time interval and calculated separately for eachdiscrete distance point By having a limited time windowfor the statistics the sensor system can adapt to changingenvironments without needs for recalibration Optionallyfor achieving a better robustness and generalization thesedistance-wise statistics can be blurred with neighboringdistance points

43 Object Boundary Detection with Density-Based StatisticalClustering (DBStaC) As a next step the standardized datagiven in the previous section is used for object detectionThetwo-dimensional data 119890119894(119899119863) is passed through a modifieddensity-based clustering algorithm based on the originalDBSCAN (Density-Based Spatial Clustering of Applicationswith Noise [34]) which is able to perform clustering onnoisy data by aggregating core points of data peaks In ourcase these data peaks can be the result of statistical outliersdue to physical objects passing by the sensor in comparisonwith the base distribution of the signal when no objects arepresent Our proposed modified algorithm in combinationwith the input preprocessing steps from (7) and (9) calleddensity-based statistical clustering (DBStaC) can be used onboth statistical hypothesis test results and in our case of asimplified algorithm data standardized based on the a prioristatistics

431 Choice of Clustering Algorithm The choice of a suitableclustering algorithm from the field of unsupervised learn-ing techniques for our application was subject to severalrequirements While many clustering algorithms can aggre-gate nearby points in a space with arbitrary dimensionality

and sparsely distributed data points it is required here toincorporate the weight of data points coming from ourthe nonsparse standardized and clipped data field 119890119894(119899119863)described before in Section 42

Furthermore the algorithm should be able to detectimportant data peaks based on a local density while forthe whole cluster no penalty for an infinite extension inthe time dimension should be given This is necessarybecause the dwell time of an object in the sensor range isarbitrary especially due to different movement speeds oftraffic participants or even in a traffic jam situation

As a third requirement an anisotropy of the clusterproperties and shape is desired as we have two dimensionsthat are time (index of different impulse responses) anddistance (specific point in the impulse response itself)which have highly different characteristics A typical sce-nario where anisotropy would not be required is 2D imageprocessing where both image dimensions have similarproperties

432 Original Definition of DBSCAN First we want toprovide a basic understanding of the original DBSCANalgorithm in general and the specific terms coming with thealgorithmic definition Then we can introduce the modifi-cations for our specific application Readers interested in thecomplete formal description are referred to [34 35] onwhichthe following original definitions are based

Core Point The general principle of DBSCAN is the analysisof a set 119863 of points in a 119896-dimensional space Then for anygiven point p isin 119863 the 120576-neighborhood of this point isevaluated meaning that all points q isin 119863 with a distancedist(p q) le 120576 belong to the neighborhood of p denoted 119873The distance metric dist(p q) is typically chosen to be the1198711 or 1198712 norm and 120576 is preset This results in the score 119899120576of a point p being the number of other points q falling intothe 120576-neighborhood Then the point p is called core point if119899120576 ge MinPts with MinPts being a preset threshold for corepoint detection All points in the 120576-neighborhood of a corepoint are part of a cluster set 119862 and they are called borderpoints if they are not core points themselves

Direct Density-Reachability A point p is directly density-reachable from a point q if q is a core point and p is in the120576-neighborhood of q This property is symmetric if p and qare a pair of core points [34]

Density-Reachability A point p is density-reachable from apoint q if a chain of 119899 points p1 p119899 with p1 = q p119899 =p exists where every point p119894+1 is directly density-reachablefrom p119894 forall119894 isin [1 sdot sdot sdot 119899 minus 1] [34]Density-Connectivity A point p is density-connected toanother point q if there is a point o such that both p and qare density-reachable from o [34]

Cluster Definition Given the set of all points 119863 a cluster119862 is a nonempty subset of 119863 which satisfies the followingconditions [34]


(1) forallp q if p isin 119862 and q is density-reachable from p thenq isin 119862 (maximality)

(2) forallp q isin 119862 p is density-connected to q (connectivity)

Any cluster 119862 is then uniquely determined by its core points

433 Modified Core Point Definition for DBStaC In ourspecial case the algorithm is modified to suit the needs forclustering our two-dimensional space therefore 119896 = 2 inorder to represent the time and distance dimension of ouracquired data Furthermore our data points are weighted bytheir standardized valueTherefore for a pointp = (119894 119899119863) theweight 119890119894(119899119863) is assigned which is not an option in the classicDBSCAN algorithm Instead of using the dist metric fordetermining the 120576-neighborhood anisotropy is introducedby choosing the 120576-neighborhood as a rectangular area withdimension (2120576119879+1)times(2120576119863+1) symmetrically placed aroundthe position (119894 119899119863) of the core point with 120576119879 and 120576119863 givenThen all points q(1198941015840 1198991015840119863) with |1198941015840 minus 119894| le 120576119879 and |1198991015840119863 minus 119899119863| le 120576119863belong to the 120576-neighborhood119873 of p = (119894 119899119863) and p is a corepoint if

119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) ge MinSum (12)

With this definition all data points of our two-dimensional data set are now taken into account not justby their count but by the sum of their assigned weightsin the specific 120576-neighborhood The implementation of thefinal clustering is omitted at this point and widely describedin literature [36] With the original algorithm implementedadaptions only have to be made in the core point definitionand neighborhood metrics

434 False Alarm Rate In the clustering process everyelement of 119890119894(119899119863) is tested for the core point propertyrequiring both compensation for multiple comparisons andan approximation of the false alarm rate for a single pointitself For the latter we will providemeasures for determiningthe false alarm rate under the approximative distributionassumptions from Section 42 With the threshold decisionfrom (12) for a core point given MinSum and the 120576-neighborhood with

119860 = (2120576119879 + 1) (2120576119863 + 1) (13)

points in the rectangular region we want to calculate the falsealarm probability

119875119891 = 119875 (119878 (119894 119899119863) ge MinSum | 119874) (14)

where119874 is the event that no relevant object reflection yieldinga core point is present meaning that all summands followthe base distribution without outliers 119878(119894 119899119863) is the sum ofweights in the 120576-neighborhood

119878 (119894 119899119863) =119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) (15)

n = 100

10 20 30 40 50 60 700x

n = 50n = 20

n = 10

n = 1n = 5

10minus5

10minus4

10minus3

10minus2

10minus1

100

Surv

ival

func

tion1minusPs(x)

Figure 7 Survival function 1 minus 119875119904(119909) for distribution 119901119904(119909) with 119899-fold convolution of 119891(119909)

In Section 42 we showed that under the assumption thatthe original envelope signal 11989010158401015840119894 (119899119863) is coming from a pro-cess with normal distribution 1198901015840119894 (119899119863) is standard-normallydistributed after standardization and 119890119894(119899119863) follows a left-censored standard normal distribution If the data points119890119894(119899119863) forall119894 119899119863 and the resulting 119878(119894 119899119863) are also hypothesizedto be statistically independent and exhibit the identical PDF119891(119909) from (10) we can write for the resulting PDF 119901119878(119909) ofthe summation of 119899 = 119860 points in (15)

119901119878 (119909) = (119891 lowast sdot sdot sdot lowast 119891)⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟119899 times

(119909) = 119891lowast119899 (119909) (16)

Please note that the statistical independence of all 119878(119894 119899119863) isnot given in the first place because the same data points inthe two-dimensional field are affected by the sliding windowin several 119894 and 119899119863 positions Then adjacent 119878(119894 119899119863) arecorrelated and a single peak in 119890119894(119899119863) could yield multiplepeaks in 119878(119894 119899119863) resulting in core points However as we seekto calculate the false alarm rate in general for this specificalgorithm adjacent core points would be clustered into asingle cluster raising only false alarm for a single object if119875119891 ≪ 1

Finally neither for the left-censored standard normaldistribution119891(119909) nor for the related half-normal distributionin general the result of 119896-fold convolution is known or givenin literature in closed form for 119896 gt 2 Therefore the finalcalculation is based on Monte Carlo simulations of theserandomprocesses For the false alarmprobability analysis weare interested in the survival function (SF) 1 minus 119875119904(119909) where119875119904(119909) is the CDF of 119901119904(119909) With 119909 = MinSum the survivalfunction yields the false alarm probability per data pointTheresults from the Monte Carlo simulations for the survivalfunction are shown in Figure 7 and were also verified withan additional false alarm Monte Carlo simulation

435 Choice of Clustering Neighborhood The neighborhoodchoice in the time direction 120576119879 affects the detection robust-ness and separability of vehicles with length 119897Veh followingeach other in a gap distance 119897Gap with speed V and passing thesensor lobe which approximately covers a section of length


Smart lighting embedded platform with Sitara SoC

Statistical object detection

Nonlinearity

Sensor platform with FPGA

Statistical test results

Object candidate boundaries

Envelope information

Time-frequency information

Statistical information

Clusterproperties and

Traffic participantobject information

Blockslicing

Time-frequency

(TF) analysis

Envelope ofcomplex

signalStandardization or

hypothesis testDensity-based

clustering

Cluster analysis

and feature

extraction

Inferencestage

Statisticsmodule

Statisticsmemory

Statistical feedback

Impulse response envelopes

Analytical signal

Controller

ADC

Waveformgenerator

DAC

Configuration

Analog frontend withdriver and duplexer

Ultrasonic sensor transducer

Timingcontrol

FilterHilbert

ℋRrarr C

metainformation

Figure 8 Signal chain with acquisition analysis and inference stages on both platforms

119897Lobe on the specific lane Given the pulse repetition rate 119879repno vehicle is covered by the sensors during the gap for 119899Gappulse times

119899Gap = (119897Gap minus 119897Lobe)(V sdot 119879rep) (17)

Also considering the detection of a single vehicle the vehicleis sampled in the lobe for 119899Veh pulses

119899Veh = (119897Veh + 119897Lobe)(V sdot 119879rep) (18)

Given 120576119879 the full neighborhood extension in time directionis 2120576119879 + 1 pulses Therefore 119899Gap ge 2120576119879 + 1 is recommendedfor a good separation of the vehicle clusters in time directionin order to have no overlapping core points Typical valuesof the time neighborhood are 0 le 120576119879 le 2 With an urbanarea example of 120576119879 = 1 119897Veh = 45m V = 40 kmh and 119897Lobe= 15m the recommended gap length becomes 119897Gap ge 483mand a vehicle is covered by 119899Veh = 54 pulse timesThe secondneighborhood parameter 120576119863 in distance direction is mostlyrelevant when it comes to multilane separation of multiplevehicles In general our approach in this paper is to learnthe typical vehicle speed and distance profile without a directcalculation but by learning the scenario for a short trainingperiod (eg one hour) and then setting the neighborhoodparameters in a parameter optimization which is describedin the following section

44 Signal Processing Concept In Figure 8 the previouslydescribed statistical object detection is integrated into thecomplete object detection concept with feature extractionand inference It allows detecting candidates for detectedobjects representing traffic participants which are then usedto gather further information from several signal processingpaths The preprocessing and main processing stages are alsodivided in hardware into the sensor platform FPGA and theSitara SoC respectively

441 Preprocessing In contrast to the sensor platform struc-ture described previously in Section 3 here the signal flowand processing are in focus As can be seen in the leftpart of Figure 8 the sensor platform is configured by themain system on the Sitara SoC and performs the wholesignal transmission and reception The preprocessing stepsof filtering Hilbert transformation and possible sample rateadjustments are performed using a dedicated architecture onthe FPGA before slicing the blocks from the different sensorsinto single impulse responses which are then transmitted tothe main processing system with the Sitara SoC

442 Object Detection and Extraction On the right side ofFigure 8 the complete object detection and feature extractionconcept is shown It is based on four main informationpaths going into the cluster analysis and feature extractionblock at the end whose output is then used for inference ofdetected traffic participants By performing a fusion of thesefour paths all necessary information for the analysis can beprovidedThe different components of the processing systemare described in the following

Statistical Object Detection with Resulting Object CandidateBoundaries The basic principle of object clustering on thestandardized data for boundary detection has already beendescribed before in Sections 42 and 43 The only addition isan optional nonlinearity after standardization or the optionalhypothesis testing (which is not considered here) The non-linearity can be modified to perform further transformationson the standardized data With the DBStaC algorithm theenvelope signal at the beginning of the statistical objectdetection block is analyzed by the statistics module togetherwith its statistics memory Only a predetermined amountof past data can be stored in this fixed-size memory forcalculating the base statistics The boundaries of resultingclusters of the algorithm are taken as boundaries for furtheranalyses in the time-distance-field of the different data pathsInside each clusterrsquos area which was determined by theobject boundary information path the cluster analysis stage


performs an analysis of the statistical envelope and time-frequency-information path

Statistical Feedback The statistical feedback path comingfrom the inference stage is important to keep the basedistribution in the statistics memory accurate to be mostlyrestricted to data without objects present This is an optionalfeature and especially helpful with high fluctuations of trafficdensity in front of the sensor as the detected objects areremoved from the statistics memory and the outlier analysisof the DBStaC-based object detection is more accurate androbust against drift

Statistical Information and Envelope Information The stan-dardized data and the original envelope signal are directlyfed into the cluster analysis stage For feature extraction andaccurate information on the object reflection characteristicsproperties like the accurate position of the objectrsquos firstreflection and the mathematical two-dimensional center ofmass of the signal or the geometry are analyzed Especiallywith a priori geometric information on the system setup ina specific scenario effects caused by the beam width of thetransducer can be exploited Even if a car moves perfectlysideways through the cone-shaped beam of the sensor itforms a typical arch-shaped pattern which could already beseen in the initial example signal Figure 6 This arc-shapedreflection pattern is also known frommarine sonar systems asldquofish arcrdquo and caused by the fact that at the point in timewhena vehicle first enters the beam the reflection path is diagonaland is not fully straight until the car is directly in front of thesensorThis can be used to support the information on vehiclesize and type

Time-Frequency Information The time-frequency informa-tion is mostly irrelevant as long as the sensor is oriented per-pendicular to the vehiclemovement on the streetHowever incase that the sensor is oriented partially sideways or a secondsensor with such orientation is available velocity informationcan be gathered by analyzing the instantaneous frequency ofthe signal for Doppler shift estimation

Inference Stage The inference stage performs the final deci-sion whether the cluster detected in the DBStaC stage origi-nated from a traffic participant in the sensor range and allowsthe estimation of further object parameters and a simpleclassification of the vehicle type using a supervised learningstage with a Support Vector Machine As our analyses in thisarticle are focused on the DBStaC object candidate detectionalgorithm itself no further rejection of objects is performedby the inference stage in this case Only objects detectedoutside of the specific lanes for example on the sidewalk arerejected

45 ImplementationDetails Themain parts of the processingchain shown on the right side of Figure 8 were implementedin C++ including optimized versions for embedded systemoperations while parts of the postprocessing and learningtechniques as part of the cluster analysis and inference stageswere implemented in Python The core processing module

can be used for both algorithmic and performance analysispurposes on general LinuxBSD systems and the embeddedLinux system running on the ARM Cortex A8 as part ofthe Sitara SoC For algorithmic evaluation and parameteroptimization in the next section capabilities of the coremodule will also be used as part of an evaluation systemThe numerical precision requirements for embedded systemoperation were relaxed from 64-bit floating point to 32-bitfloating point largelywithout compromising performance inorder to exploit the NEON floating point accelerator of theSitara SoC For the processing chain with the DBStaC algo-rithm real-time operation is then possible for two attachedultrasonic sensors This is the case for typical parameter setsas the algorithmic runtime of DBSCAN is not fixed and islargely dependent on the choice of parameters (size of 120576-neighborhood and decision thresholds) and traffic situationyielding different cluster sizes Edge cases like extremely largeclusters for example due to a traffic jam situation with verylong dwell times of an object in the sensor range have to betaken care of in the practical implementations

For the sensor platform with the FPGA shown on theleft side of Figure 8 dedicated accelerators for filtering andprocessing are used together with a MicroBlaze soft core forthe coordination of waveform transmission and data acquisi-tion The configuration is controlled by the main processingsystem side with the Sitara SoC A timing reference signalis acquired using GPS together with the high-precision PPSreference which allows global synchronization of multiplesensors attached to multiple platforms with an optimizationof the transmission and interference patterns Exchange ofresulting information for example for sensor fusion anddistributed processing for trajectory calculation of objectspassing by multiple systems is possible on both the level ofa local wireless meshing between the systems and a globalcommunication using cellular networks

5 Evaluation Scenarios and Methodology

Evaluation of the complete system concept for traffic partici-pant detection requires analyses both in real-world scenariosand in isolated conditions with synthetic scenarios suchas on automotive test tracks A variety of European-widetest measurements have been performed with the systemdescribed in Section 52 whichwill be analyzed anddiscussedin Section 6 For the reference data acquisition a videocamera is running in parallel which is afterwards used tolabel annotate and classify traffic participants on the timeline These labels can then be utilized to evaluate the perfor-mance in terms of several metrics for quantifying detectionperformance parameter estimation and classification Ingeneral it is desirable to have specific correctness informationfor every object instead of just taking the overall numbersof detected objects for the analysis which is supportedby an exact pairing of detected objects and reference datain the following Then parameter sets for the system areoptimizedwith regard to different performancemetrics usinga central (hyper)parameter optimizer as part of an evaluationframework which is described in the following This processcan also be deployed to a high-performance cluster (HPC)


Sensor and reference data

Recordedsensor data

Recordedreference video

Multilane video labeling and object annotation tool

mySQL Reference object information database

Deployable parameter point evaluation process (using hyperopt-worker) Hyperparameter optimization and evaluation coordinator

mongoDB evaluation and parameter space database

Coordination

Evaluation

Processing

Algorithmic baseconfiguration

Hyperopt-server (Python)Centrally coordinatedhyper-parameter space exploration andoptimization

Hyperparameter and configuration spacedefinition

Currentconfiguration set

Sensor dataprocessing chainMain C++-based algorithmic processing chain module(also suitable for embedded system operation)

PostprocessingPython-based parameter estimation and extraction of further characteristics

Scoring based on comparison with reference dataPython based parameter estimation and extraction of further characteristics

Results optimization loss parameter errors

Figure 9 Evaluation framework and methodology for parameter space exploration

51 Parameter Space Exploration and Optimization Method-ology In the following the framework for parameter spaceexploration of the whole system with its correspondingalgorithms is specified The basic structure is shown inFigure 9

Parameter Set Evaluation Concept As a main principle(hyper)parameter sets are given to aworker shown in the cen-tral block in the diagram Each worker then evaluates a singleparameter set (deployed with the Coordination subsystem)for real-world recorded sensor data using the C++-basedmain processing chain and Python-based postprocessing aspart of the Processing subsystem The processing principleswere described in the implementation description of thisarticle in Section 45 Then in the Evaluation subsystemthe objects resulting from the processing stage are takentogether with data from an object reference database Thisallows the analysis of several performance metrics such asthe detection 1198651 score and the correct lane assignment whichyields a performance score or optimization loss as a qualityindicator of the parameter set This is then reported back tothe Coordination subsystem of the worker

ReferenceDatabase On the left side of the diagram the sensorand reference data are shown Reference object data is storedin amySQLdatabase A reference video is recorded in parallelto a sensor data recording This video is then manuallyprocessed using a labeling tool which is also shown in thediagramThe tool allows labeling traffic participant objects onmultiple lanes in a timeline view and adding annotations tothese objects such as the vehicle type vehicle size movementdirection and velocity which is stored in the database Byanalyzing the raw video material itself the tool also providesthe activity level in the video material which supportsfinding the next vehicle passing by in less frequented timeranges

Cluster-Label-Pairing for Scoring Both the resulting clustersfrom the processing of the raw sensor data and the objectsin the reference database are events extended in the timedimension with a specific start and end time In the detectionperformance analysis of the system each cluster (in caseof perfect detection) would be paired with the correspond-ing reference label object As the exact cluster-label-pairassignment is not the case in a real scenario there can beambiguities or false-positivefalse-negative cases Thereforea bipartite cluster-label-graph is built up from the clusterand reference data with edges between the nodes of specificcluster-label candidate pairs Then a maximum cardinalitybipartite matching is performed With the resulting cluster-label-pairs and the known sets of all reference and clusterobjects all scorings such as precision recall accuracy and1198651-score can be calculated and reported back for the specificparameter set

Coordination of (Hyper)parameter Sets Given the possibilityof using a worker process to evaluate parameter sets with thedescribed procedure the exploration of the (hyper)parameterspace needs to be coordinated and optimizedThis is achievedby using the hyperopt Python package [37] which allowsthe definition of several (hyper)parameters with their accord-ing distributions and properties and then to automaticallyexplore this space using simulated annealing random searchand tree of Parzen estimator techniques for optimizationThe parameter sets and their according loss results are storedin a MongoDB database New parameter sets are integratedinto a base configuration file which is then assigned to theworker process In order to speed up the optimization andtraining procedure a large number of workers are deployedon an HPC Typically good convergence is achieved after10000 optimization iterations If a faster convergence isdesired starting points for the parameters can be calcu-lated with the investigations done in Section 43 Typical


Table 2 Overview of test scenarios and their characteristics

Scenarioname

Number oflanes

Sensordistancerange1

Sensormountingheight

Typical speedrange

Pulse interval119879rep

Description

Urban1 2 75m 35m 20ndash50 kmh 100ms

Urban alley street with trees on both sides andparked cars on the adjacent side mixed use bycars buses and bicyclists sensor placement in ahorizontal distance of 1m to curb


Sensor placement behind sidewalk distance tocurb 2m mixed use by cars buses andbicyclists reflective trash containers onadjacent side

Urban3 2 7m 5m 20ndash50 kmh 100msTypical ldquourban canyonrdquo with large buildings onboth side close to the street and narrowsidewalks distance to curb 20 cm

TestTrack - gt15m 35m 5ndash100 kmh 50ndash100ms

Automotive test track without reflectiveobjects only asphalted road surface in sensorrange low rate of vehicles passing by withhighly different speeds

1The distance range is the distance from the ground projection point of the sensor position to the curb side or delimiter of the adjacent street side For thecalculation of the time-of-flight equivalent distance both this sensor distance range and the mounting height have to be incorporated

(hyper)parameters which are part of the optimization pro-cess are as follows

(i) Dimensions of 120576-environment (120576119879 120576119863) for theDBStaCalgorithm and the core point threshold level MinSum(see Section 433)

(ii) Properties of the statistics memory such as the statis-tics memory size (past information see Section 442)and a storage subsampling factor

(iii) Optional use of statistical feedback (see Section 442)in the scenario for base distribution stabilization

(iv) Optional clipping threshold for calculated statisticalnorms in standardization to stabilize against outliers

(v) Optional use of matched filtering for the receivedenvelope signal

52 Test Scenarios In Table 2 the different test scenarios aregiven with their characteristics and description covering avariety of urban environments and synthetic measurementson an automotive test track The scenarios are identified by aspecific name and will be referenced in the following whenperformance results are given

6 Results and Discussion

Results of the field tests and evaluations are given in thischapter First the according performance metrics are intro-duced and specified followed by the results for the differentscenarios and finally a discussion of the results

61 Evaluation Metrics In this paper the main focus lieson the detection of traffic participants as objects and theircorrect assignment to the lanes of the street For describingperformance in our object detection scenario no calculation

of the true-negative count is possible as we can have anarbitrary number of objects in our data timeline Thereforethe widely used 1198651 score is chosen as the main performancemetric for detection being the harmonic mean of precisionand recall The precision 119875 is defined as the number of truepositives (119879119875) divided by all positive outcomes

119875 = 119879119875119879119875 + 119865119875 (19)

In contrast the recall 119877 is defined as the number of truepositives divided by the sum of true positive and false-negative (119865119873) outcomes

119875 = 119879119875119879119875 + 119865119873 (20)

Then the 1198651 score is defined as follows

1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)

A further advantage is that the combination of precisionand recall into the 1198651 score allows being independent ofrequirements biases which can be a focus on either precisionor recall performance for example whenmore false positives(119865119875) or negatives are acceptable in a specific scenario1198651 scorecalculation only requires the information of true positivesfalse positives and false negatives These numbers are cal-culated based on the bipartite cluster-label-graph assignmentintroduced in Section 51

The second important metric is the assignment of objectsto the correct lanes for determining the direction of the trafficflowWithmultiple lanes as amulticlass problem the1198651 scorecannot be used directly As an alternative the weighted 1198651score is used which first calculates the 1198651 score for every lane(correct or incorrect assignment to specific lane) and thencalculates a weightedmean of all lane scores with the numberof true reference object instances as the weight


Table 3 Field test evaluation results for different scenarios

Scenarioname Object count 1198651 score

detectionWeighted 1198651 scorelane assignment

Urban1 224 098 091Urban2 133 092 095Urban3 305 097 098TestTrack 17 094 minus

62 Discussion of Test Results The performance evaluationresults for the different scenario with optimized parametersare shown in Table 3 Both detection and lane assignmentscores are always larger than 09 being a very good per-formance level in real scenarios and satisfying the initiallystated high requirements on traffic information quality Thesingle-sensor ultrasonic traffic participant detection is there-fore possible with high performance without exploiting anydirectional information of the signal This is facilitated bythe algorithms proposed in this paper which solved severalprevailing problems in ultrasonic sensing techniques

The multilane assignment performance is also very highand of special importance As no directional or velocity infor-mation is availablewith these sensors the known informationof a fixed lanersquos driving direction yields the final traffic flowinformation with direction A possible problem is showingup in the scenario Urban1 with the lowest weighted 1198651 scorefor assignment of 091 people were driving in the middle ofthe two-lane street several times which compromised laneassignment performance

Further future long-term analyses are required to analyzethe long-term stability in scenarios with a highly fluctu-ating traffic flow and day-and-night cycle However withthe system structure which relies on the standardized datatogether with the statistics memory the sensitivity to long-term sensor drift and also production variations is limitedFurther investigation is also required on the impact of thestatistical feedback for stabilization of the statistical baseinformation As the parameter optimization is based on onlya few parameters the susceptibility to overfitting effects islimited in the shown results Still for future investigationsalso the test performancewith regard to futuremeasurementsin the same scenario is of high interest

7 Conclusion and Future Work

In this paper it was shown that sidefire ultrasonic sensingis a viable option for multilane traffic participant sensinggiving very good performance results in real-world urbanscenarios with a single sensor The functionality is enabledby a novel combination of standardization techniques basedon windowed statistics and a modified density-based clus-tering algorithm together called DBStaC Several evalua-tions were performed both in real urban environmentsand for special cases on an automotive test track Theproposed system is integrated into an evaluation and param-eter optimization methodology whose reference system isflexible to be extended with further object characteristics infuture

The ultrasonic sensor system deployed together withthe street lamp platform was presented to be a Smart Citytechnology suitable for high-coverage deployment in citiesenabling traffic monitoring and further applications Withthis platform distributed processing and sensor fusion canexpand the possibilities of using available traffic participantinformation even more for example for future applicationslike trajectory prediction Further future work can be theuse of ultrasonic sensor arrays for acquiring directional andvelocity information In the field of algorithmic improve-ments the investigation of more advanced hypothesis test-ing techniques and channel statistics analyses could yieldfurther performance improvements Also the comparisonof the presented algorithmic concept with state-of-the artsupervised machine learning algorithms such as recurrentneural networks is planned

Conflicts of Interest

The authors declare that they have no conflicts of interest

References

[1] H van Essen and A van Grinsven ldquoFinal report appendix 9Interaction of GHG policy for transport with congestion andaccessibility policiesrdquo Project Report EU Transport GHG 20502012

[2] ldquoRoadmap to a single european transport area - towards acompetitive and resource efficient transport systemrdquo WhitePaper European Commission 2011

[3] H Mayer ldquoAir pollution in citiesrdquo Atmospheric Environmentvol 33 no 24-25 pp 4029ndash4037 1999

[4] R Arnott A de Palma and R Lindsey ldquoDoes providing infor-mation to drivers reduce traffic congestionrdquo TransportationResearch Part A General vol 25 no 5 pp 309ndash318 1991

[5] A Zanella N Bui A P Castellani L Vangelista and M ZorzildquoInternet of things for smart citiesrdquo IEEE Internet of ThingsJournal vol 1 no 1 pp 22ndash32 2014

[6] N Buch S A Velastin and J Orwell ldquoA review of computervision techniques for the analysis of urban trafficrdquo IEEE Trans-actions on Intelligent Transportation Systems vol 12 no 3 pp920ndash939 2011

[7] S Himmel M Ziefle and K Arning From Living Space toUrban Quarter Acceptance of ICT Monitoring Solutions in anAgeing Society Springer Berlin Germany 2013

[8] W Xue L Wang and D Wang ldquoA prototype integratedmonitoring system for pavement and traffic based on anembedded sensing networkrdquo IEEE Transactions on IntelligentTransportation Systems vol 16 no 3 pp 1380ndash1390 2015

[9] J Gajda R Sroka M Stencel A Wajda and T Zeglen ldquoAvehicle classification based on inductive loop detectorsrdquo inProceedings of the 18th IEEE Instrumentation and MeasurementTechnology Conference RediscoveringMeasurement in the Age ofInformatics pp 460ndash464 May 2001

[10] L Klein D Gibson and M Mills Traffic Detector Handbookvol 1 no FHWA-HRT-06-108 Federal Highway Administra-tion Turner-Fairbank Highway Research Center 3rd edition2006


[11] M Hallenbeck and H Weinblatt Equipment for CollectingTraffic Load Data Transportation Research Board NCHRPReport 509 2004

[12] N Garber and L Hoel Traffic and Highway Engineering SIEdition Cengage Learning 2014

[13] R Mobus and U Kolbe ldquoMulti-target multi-object trackingsensor fusion of radar and infraredrdquo in Proceedings of the IEEEIntelligent Vehicles Symposium pp 732ndash737 IEEE June 2004

[14] I Urazghildiiev R Ragnarsson P Ridderstrom et al ldquoVehicleclassification based on the radar measurement of height pro-filesrdquo IEEE Transactions on Intelligent Transportation Systemsvol 8 no 2 pp 245ndash253 2007

[15] I Urazghildiiev R Ragnarsson K Wallin A Rydberg PRidderstrom and E Ojefors ldquoA vehicle classification systembased on microwave radar measurement of height profilesrdquoin Proceedings of the 2002 International Radar Conference pp409ndash413 Edinburgh UK October 2002

[16] J Tao S Chen L Yang and Y Hu ldquoUltrasonic techniquebased on neural networks in vehicle modulation recognitionrdquoin Proceedings of the The 7th International IEEE Conference onIntelligent Transportation Systems pp 201ndash204 October 2004

[17] E U Warriach and C Claudel ldquoPoster abstract A machinelearning approach for vehicle classification using passiveinfrared and ultrasonic sensorsrdquo in Proceedings of the 12thInternational Conference on Information Processing in SensorNetworks IPSN 2013 - Part of CPSWeek 2013 pp 333-334 April2013

[18] T Matsuo Y Kaneko and M Matano ldquoIntroduction ofintelligent vehicle detection sensorsrdquo in Proceedings of the1999 Intelligent Transportation Systems Conference pp 709ndash713Tokyo Japan

[19] H Kim J-H Lee S-W Kim J-I Ko and D Cho ldquoUltrasonicVehicle Detector for Side-Fire Implementation and ExtensiveResults Including Harsh Conditionsrdquo IEEE Transactions onIntelligent Transportation Systems vol 2 no 3 pp 127ndash134 2001

[20] V Agarwal N V Murali and C Chandramouli ldquoA cost-effective ultrasonic sensor-based driver-assistance system forcongested traffic conditionsrdquo IEEE Transactions on IntelligentTransportation Systems vol 10 no 3 pp 486ndash498 2009

[21] C Heipke ldquoCrowdsourcing geospatial datardquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 65 no 6 pp 550ndash5572010

[22] S Hind and A Gekker ldquoOutsmarting traffic together Drivingas social navigationrdquo Exchanges the Warwick Research Journalvol 1 no 2 pp 165ndash180 2014

[23] M B Sinai N Partush S Yadid and E Yahav ldquoExploiting socialnavigationrdquo httpsarxivorgabs14100151v1

[24] R Bauza J Gozalvez and J Sanchez-Soriano ldquoRoad trafficcongestion detection through cooperative Vehicle-to-Vehiclecommunicationsrdquo in Proceedings of the 35th Annual IEEEConference on Local Computer Networks LCN 2010 pp 606ndash612 October 2010

[25] S Jeon E Kwon and I Jung ldquoTrafficmeasurement on multipledrive lanes with wireless ultrasonic sensorsrdquo Sensors vol 14 no12 pp 22891ndash22906 2014

[26] Y Jo and I Jung ldquoAnalysis of vehicle detection withWSN-basedultrasonic sensorsrdquo Sensors vol 14 no 8 pp 14050ndash14069 2014

[27] H Kuttruff Room Acoustics Spon Press New York NY USA5th edition 2009

[28] A V Oppenheim R W Schafer and J R Buck Discrete-TimeSignal Processing Prentice-Hall Inc Upper Saddle River NJUSA 2nd edition 1999

[29] F J Harris ldquoOn the use of windows for harmonic analysis withthe discrete Fourier transformrdquo Proceedings of the IEEE vol 66no 1 pp 51ndash83 1978

[30] B Gold A V Oppenheim and C M Rader ldquoTheory andimplementation of the discrete Hilbert transformrdquo Symposiumon Computer Processing in Communications vol 19 1970

[31] M Meissner ldquoAccuracy issues of discrete hubert transform inidentification of instantaneous parameters of vibration signalsrdquoActa Physica Polonica A vol 121 no 1A pp A164ndashA167 2012

[32] L Xu and T Xu Digital Underwater Acoustic CommunicationsElsevier Science 2016

[33] S C Kumbhakar and C A K Lovell Stochastic FrontierAnalysis Cambridge University Press Cambridge UK 2003

[34] M Ester H-P Kriegel J Sander and X Xu ldquoA density-basedalgorithm for discovering clusters a density-based algorithm fordiscovering clusters in large spatial databases with noiserdquo inProceedings of the Second International Conference onKnowledgeDiscovery and Data Mining ser KDDrsquo96 pp 226ndash231 AAAIPress 1996

[35] J Gan and Y Tao ldquoDBSCAN revisited Mis-claim un-fixabilityand approximationrdquo in Proceedings of the 2015 ACM SIGMODInternational Conference on Management of Data ser SIG-MODrsquo15 pp 519ndash530 ACM New York NY USA 2015

[36] E Schubert J Sander M Ester H P Kriegel and XXu ldquoDBSCAN Revisited Revisitedrdquo ACM Transactions onDatabase Systems (TODS) vol 42 no 3 pp 1ndash21 2017

[37] J Bergstra D Yamins and D D Cox ldquoHyperopt A Pythonlibrary for optimizing the hyperparameters ofmachine learningalgorithmsrdquo in Proceedings of the 12th Python in Science Confer-ence S van der Walt J Millman and K Huff Eds 2013

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018


Active and Passive Electronic Components

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of



Journal ofEngineeringVolume 2018

SensorsJournal of



RotatingMachinery


Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom


if they are distributed sparsely and do not achieve sufficientcoverage

In general two main types of techniques achieving morecomprehensive traffic information have been establishedin the last years A first category is the fixed or mobileinfrastructural distribution of sensors like inductive loopsand traffic camerasThe second type is represented by systemsrelying on end-user based distributed information for exam-ple movement data from smartphone users or informationgathered using V2X communication approaches Details andcharacteristics of several techniques are discussed in thispaper considering not only metrics like traffic informationquality and performance in different scenarios but also pos-sible privacy issues which may arise with certain techniques

The novel ultrasound-based traffic sensing techniquewhich is the focus of this paper belongs to the category ofinfrastructural sensing A significant difference to existingultrasonic sensor systems is the placement of the sensors ina so-called sidefire setup This means that the sensors aremounted in a height of at least threemeters and face the streetsideways with downtilted sensors while being positionedon the side of the street Previously existing systems weremainly single-lane ultrasonic sensors measuring the heightprofile top-down for example from a special sensor gantryabove the street with a simple first-reflection processingThe monitoring of multiple lanes is possible with our setupwhile at the same time the requirements in terms of algorith-mic evaluation and processing are increased in comparisonto these previously existing systems However with theseimprovements our approach allows operation in a highlyreflective acoustic environment which is present in urbanareas and permits simply mounting the sensor on the sideof the street This is enabled by a new approach with acombination of statistical signal processing clustering andinference algorithms for traffic participant object detection

This sidefire ultrasonic traffic sensing technique is partof a development which focuses on intelligent infrastructuresolutions more specifically on intelligent street lights Anexemplary integrated setup is shown in Figure 1 Theselight modules can be used to replace inner components ofstreet lamps for example by retrofitting the widespread typeof high-pressure sodium-vapor lamps Then they can actas a flexible platform for many applications by employingthe processing communication and light control actuationcapabilitiesWith the transition to LED-based lighting whichis triggered by high energy costs and several EU directives[5] a large proportion of the existing street lamps will haveto be replaced in the next decade offering potential for anintegration of Smart City solutions The seamless integrationinto infrastructure allows an almost complete coverage ofrelevant urban traffic to provide the required coverage densityof traffic information Furthermore the major problem ofpower supply for sensing devices is alleviated as the lamps areconnected to the mains power which eliminates the need forenergy-constrained and expensive battery-powered devices

The rest of this article is organized as follows Section 2provides the related work of urban traffic sensing techniquesin general and ultrasonic sensing in specific Section 3describes the system architecture together with the specific

Figure 1 Integrated street lamp head with a single ultrasonic sensorhead

scenario and typical sensor setup Section 4 provides thealgorithmic approach to the sensor signal processing systemand object detection inference In Section 5 the evaluationmethodology and framework for reference data are pre-sented Section 6 gives results of real-world measurementsand provides an evaluation from the perspective of detectionand classification together with a state-of-the-art technologycomparison Section 7 then concludes the article

2 Related Work

Ultrasonic traffic participant detection falls into the field oftraffic sensing with a large number of techniques availablein research and in practical application In this sectionwe therefore provide a state-of-the-art overview from twoperspectives First we give a comprehensive overview oftraffic monitoring techniques in general with a discussionof their performance strengths and weaknesses in order toshow the technological gaps and potential In the second partspecific technical approaches in ultrasonic traffic sensingwhich are rare both in concepts and in practical implemen-tations are given This application-centric perspective showsthe drawbacks and unresolved problems in current state ofthe art

21 Traffic Monitoring Techniques from a Broad View In theintroductory section two main categories of traffic informa-tion acquisition systems were introduced These are on theone hand infrastructure-based sensing systems working ina temporarily or permanently fixed location On the otherhand the type of end-user assisted sensing systems incor-porates distributed information acquired from the driverstheir technical devices and the vehiclersquos internal sensorsitself In the following a state-of-the-art overview togetherwith a technological comparison is given for nowadays mostrelevant traffic monitoring techniques of both categoriesA condensed overview of the properties is also given inTable 1

Themostwidely used approach to traffic sensing is the useof infrastructure-based solutions which represents a mainlycentralized approach In general authority of the data iskept with the operator and thus several of the problemscoming with user-centric techniques do not arise Howeverthe sensing systems have to be deployed over the whole cityor at least at critical locations in terms of traffic Achieving therequired density for an extensive urban traffic monitoring istherefore paramount which brings along high effort and costas a centralized solution






































Θ









119904119879 (119905) =infinsum119899=0

ℎ119899 (119905 minus 119899 sdot 119879rep) (1)








(PowerTxRxControl)


Ethernet + PoE



ADC


andacquisition

Signalpreprocessing


Controland

managementprocessor

DDR3 memory

PoE power supply

Industrial Ethernet

PHY


ADC

DAC



0

Tran

smit

signa

l am

plitu

de




TＱＨ＞







119904119894 (119905119863) = 119904 (119894 sdot 119879rep + 119905119863) (3)


119909119894 (119899119863) = 119904 (119894 sdot 119873rep119879 + 119899119863 sdot 119879) (4)




123456789

10

Dist

ance

dim

ensio

n

10

20

30

40

50

60

Dist

ance

dim

ensio

n




impu

lse d

elay

tim

e (m

s)

equi

vale

nt d

istan

ce (m

)





11989010158401015840119894 (119899119863) = 1003816100381610038161003816119909119894 (119899119863)1003816100381610038161003816 (6)






1198901015840119894 (119899119863) = 11989010158401015840119894 (119899119863) minus 120583119899119863 (119894)120590119899119863 (119894) (7)




0 le 119899119863 le 119873rep minus 1 119894 ge 119873119879119908

(8)





119890119894 (119899119863) = max (1198901015840119894 (119899119863) 0) (9)


119891 (119909) =

1radic2120587119890minus11990922 + 1

2120575 (119909) 119909 ge 00 otherwise (10)


119865 (119909) =

12erf (

119909radic2) + 1


(11)


















119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) ge MinSum (12)



119860 = (2120576119879 + 1) (2120576119863 + 1) (13)


119875119891 = 119875 (119878 (119894 119899119863) ge MinSum | 119874) (14)


119878 (119894 119899119863) =119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) (15)

n = 100

10 20 30 40 50 60 700x

n = 50n = 20

n = 10

n = 1n = 5

10minus5

10minus4

10minus3

10minus2

10minus1

100

Surv

ival

func

tion1minusPs(x)




(119909) = 119891lowast119899 (119909) (16)







Nonlinearity









Blockslicing

Time-frequency

(TF) analysis

Envelope ofcomplex



clustering

Cluster analysis

and feature

extraction

Inferencestage

Statisticsmodule

Statisticsmemory



Analytical signal

Controller

ADC

Waveformgenerator

DAC

Configuration



Timingcontrol

FilterHilbert

ℋRrarr C

metainformation
























Recordedsensor data






Coordination

Evaluation

Processing

















Scenarioname

Number oflanes



Typical speedrange


Description




















119875 = 119879119875119879119875 + 119865119875 (19)


119875 = 119879119875119879119875 + 119865119873 (20)


1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)
















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia







































Θ









119904119879 (119905) =infinsum119899=0

ℎ119899 (119905 minus 119899 sdot 119879rep) (1)








(PowerTxRxControl)


Ethernet + PoE



ADC


andacquisition

Signalpreprocessing


Controland

managementprocessor

DDR3 memory

PoE power supply

Industrial Ethernet

PHY


ADC

DAC



0

Tran

smit

signa

l am

plitu

de




TＱＨ＞







119904119894 (119905119863) = 119904 (119894 sdot 119879rep + 119905119863) (3)


119909119894 (119899119863) = 119904 (119894 sdot 119873rep119879 + 119899119863 sdot 119879) (4)




123456789

10

Dist

ance

dim

ensio

n

10

20

30

40

50

60

Dist

ance

dim

ensio

n




impu

lse d

elay

tim

e (m

s)

equi

vale

nt d

istan

ce (m

)





11989010158401015840119894 (119899119863) = 1003816100381610038161003816119909119894 (119899119863)1003816100381610038161003816 (6)






1198901015840119894 (119899119863) = 11989010158401015840119894 (119899119863) minus 120583119899119863 (119894)120590119899119863 (119894) (7)




0 le 119899119863 le 119873rep minus 1 119894 ge 119873119879119908

(8)





119890119894 (119899119863) = max (1198901015840119894 (119899119863) 0) (9)


119891 (119909) =

1radic2120587119890minus11990922 + 1

2120575 (119909) 119909 ge 00 otherwise (10)


119865 (119909) =

12erf (

119909radic2) + 1


(11)


















119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) ge MinSum (12)



119860 = (2120576119879 + 1) (2120576119863 + 1) (13)


119875119891 = 119875 (119878 (119894 119899119863) ge MinSum | 119874) (14)


119878 (119894 119899119863) =119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) (15)

n = 100

10 20 30 40 50 60 700x

n = 50n = 20

n = 10

n = 1n = 5

10minus5

10minus4

10minus3

10minus2

10minus1

100

Surv

ival

func

tion1minusPs(x)




(119909) = 119891lowast119899 (119909) (16)







Nonlinearity









Blockslicing

Time-frequency

(TF) analysis

Envelope ofcomplex



clustering

Cluster analysis

and feature

extraction

Inferencestage

Statisticsmodule

Statisticsmemory



Analytical signal

Controller

ADC

Waveformgenerator

DAC

Configuration



Timingcontrol

FilterHilbert

ℋRrarr C

metainformation
























Recordedsensor data






Coordination

Evaluation

Processing

















Scenarioname

Number oflanes



Typical speedrange


Description




















119875 = 119879119875119879119875 + 119865119875 (19)


119875 = 119879119875119879119875 + 119865119873 (20)


1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)
















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




(PowerTxRxControl)


Ethernet + PoE



ADC


andacquisition

Signalpreprocessing


Controland

managementprocessor

DDR3 memory

PoE power supply

Industrial Ethernet

PHY


ADC

DAC



0

Tran

smit

signa

l am

plitu

de




TＱＨ＞







119904119894 (119905119863) = 119904 (119894 sdot 119879rep + 119905119863) (3)


119909119894 (119899119863) = 119904 (119894 sdot 119873rep119879 + 119899119863 sdot 119879) (4)




123456789

10

Dist

ance

dim

ensio

n

10

20

30

40

50

60

Dist

ance

dim

ensio

n




impu

lse d

elay

tim

e (m

s)

equi

vale

nt d

istan

ce (m

)





11989010158401015840119894 (119899119863) = 1003816100381610038161003816119909119894 (119899119863)1003816100381610038161003816 (6)






1198901015840119894 (119899119863) = 11989010158401015840119894 (119899119863) minus 120583119899119863 (119894)120590119899119863 (119894) (7)




0 le 119899119863 le 119873rep minus 1 119894 ge 119873119879119908

(8)





119890119894 (119899119863) = max (1198901015840119894 (119899119863) 0) (9)


119891 (119909) =

1radic2120587119890minus11990922 + 1

2120575 (119909) 119909 ge 00 otherwise (10)


119865 (119909) =

12erf (

119909radic2) + 1


(11)


















119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) ge MinSum (12)



119860 = (2120576119879 + 1) (2120576119863 + 1) (13)


119875119891 = 119875 (119878 (119894 119899119863) ge MinSum | 119874) (14)


119878 (119894 119899119863) =119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) (15)

n = 100

10 20 30 40 50 60 700x

n = 50n = 20

n = 10

n = 1n = 5

10minus5

10minus4

10minus3

10minus2

10minus1

100

Surv

ival

func

tion1minusPs(x)




(119909) = 119891lowast119899 (119909) (16)







Nonlinearity









Blockslicing

Time-frequency

(TF) analysis

Envelope ofcomplex



clustering

Cluster analysis

and feature

extraction

Inferencestage

Statisticsmodule

Statisticsmemory



Analytical signal

Controller

ADC

Waveformgenerator

DAC

Configuration



Timingcontrol

FilterHilbert

ℋRrarr C

metainformation
























Recordedsensor data






Coordination

Evaluation

Processing

















Scenarioname

Number oflanes



Typical speedrange


Description




















119875 = 119879119875119879119875 + 119865119875 (19)


119875 = 119879119875119879119875 + 119865119873 (20)


1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)
















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



123456789

10

Dist

ance

dim

ensio

n

10

20

30

40

50

60

Dist

ance

dim

ensio

n




impu

lse d

elay

tim

e (m

s)

equi

vale

nt d

istan

ce (m

)





11989010158401015840119894 (119899119863) = 1003816100381610038161003816119909119894 (119899119863)1003816100381610038161003816 (6)






1198901015840119894 (119899119863) = 11989010158401015840119894 (119899119863) minus 120583119899119863 (119894)120590119899119863 (119894) (7)




0 le 119899119863 le 119873rep minus 1 119894 ge 119873119879119908

(8)





119890119894 (119899119863) = max (1198901015840119894 (119899119863) 0) (9)


119891 (119909) =

1radic2120587119890minus11990922 + 1

2120575 (119909) 119909 ge 00 otherwise (10)


119865 (119909) =

12erf (

119909radic2) + 1


(11)


















119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) ge MinSum (12)



119860 = (2120576119879 + 1) (2120576119863 + 1) (13)


119875119891 = 119875 (119878 (119894 119899119863) ge MinSum | 119874) (14)


119878 (119894 119899119863) =119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) (15)

n = 100

10 20 30 40 50 60 700x

n = 50n = 20

n = 10

n = 1n = 5

10minus5

10minus4

10minus3

10minus2

10minus1

100

Surv

ival

func

tion1minusPs(x)




(119909) = 119891lowast119899 (119909) (16)







Nonlinearity









Blockslicing

Time-frequency

(TF) analysis

Envelope ofcomplex



clustering

Cluster analysis

and feature

extraction

Inferencestage

Statisticsmodule

Statisticsmemory



Analytical signal

Controller

ADC

Waveformgenerator

DAC

Configuration



Timingcontrol

FilterHilbert

ℋRrarr C

metainformation
























Recordedsensor data






Coordination

Evaluation

Processing

















Scenarioname

Number oflanes



Typical speedrange


Description




















119875 = 119879119875119879119875 + 119865119875 (19)


119875 = 119879119875119879119875 + 119865119873 (20)


1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)
















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




119890119894 (119899119863) = max (1198901015840119894 (119899119863) 0) (9)


119891 (119909) =

1radic2120587119890minus11990922 + 1

2120575 (119909) 119909 ge 00 otherwise (10)


119865 (119909) =

12erf (

119909radic2) + 1


(11)


















119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) ge MinSum (12)



119860 = (2120576119879 + 1) (2120576119863 + 1) (13)


119875119891 = 119875 (119878 (119894 119899119863) ge MinSum | 119874) (14)


119878 (119894 119899119863) =119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) (15)

n = 100

10 20 30 40 50 60 700x

n = 50n = 20

n = 10

n = 1n = 5

10minus5

10minus4

10minus3

10minus2

10minus1

100

Surv

ival

func

tion1minusPs(x)




(119909) = 119891lowast119899 (119909) (16)







Nonlinearity









Blockslicing

Time-frequency

(TF) analysis

Envelope ofcomplex



clustering

Cluster analysis

and feature

extraction

Inferencestage

Statisticsmodule

Statisticsmemory



Analytical signal

Controller

ADC

Waveformgenerator

DAC

Configuration



Timingcontrol

FilterHilbert

ℋRrarr C

metainformation
























Recordedsensor data






Coordination

Evaluation

Processing

















Scenarioname

Number oflanes



Typical speedrange


Description




















119875 = 119879119875119879119875 + 119865119875 (19)


119875 = 119879119875119879119875 + 119865119873 (20)


1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)
















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia







119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) ge MinSum (12)



119860 = (2120576119879 + 1) (2120576119863 + 1) (13)


119875119891 = 119875 (119878 (119894 119899119863) ge MinSum | 119874) (14)


119878 (119894 119899119863) =119894+120576119879sum119905=119894minus120576119879

119899119863+120576119863sum119889=119899119863minus120576119863

119890119905 (119889) (15)

n = 100

10 20 30 40 50 60 700x

n = 50n = 20

n = 10

n = 1n = 5

10minus5

10minus4

10minus3

10minus2

10minus1

100

Surv

ival

func

tion1minusPs(x)




(119909) = 119891lowast119899 (119909) (16)







Nonlinearity









Blockslicing

Time-frequency

(TF) analysis

Envelope ofcomplex



clustering

Cluster analysis

and feature

extraction

Inferencestage

Statisticsmodule

Statisticsmemory



Analytical signal

Controller

ADC

Waveformgenerator

DAC

Configuration



Timingcontrol

FilterHilbert

ℋRrarr C

metainformation
























Recordedsensor data






Coordination

Evaluation

Processing

















Scenarioname

Number oflanes



Typical speedrange


Description




















119875 = 119879119875119879119875 + 119865119875 (19)


119875 = 119879119875119879119875 + 119865119873 (20)


1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)
















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





Nonlinearity









Blockslicing

Time-frequency

(TF) analysis

Envelope ofcomplex



clustering

Cluster analysis

and feature

extraction

Inferencestage

Statisticsmodule

Statisticsmemory



Analytical signal

Controller

ADC

Waveformgenerator

DAC

Configuration



Timingcontrol

FilterHilbert

ℋRrarr C

metainformation
























Recordedsensor data






Coordination

Evaluation

Processing

















Scenarioname

Number oflanes



Typical speedrange


Description




















119875 = 119879119875119879119875 + 119865119875 (19)


119875 = 119879119875119879119875 + 119865119873 (20)


1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)
















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia















Recordedsensor data






Coordination

Evaluation

Processing

















Scenarioname

Number oflanes



Typical speedrange


Description




















119875 = 119879119875119879119875 + 119865119875 (19)


119875 = 119879119875119879119875 + 119865119873 (20)


1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)
















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




Scenarioname

Number oflanes



Typical speedrange


Description




















119875 = 119879119875119879119875 + 119865119875 (19)


119875 = 119879119875119879119875 + 119865119873 (20)


1198651 = 2 119875 sdot 119877119875 + 119877 = 2 sdot 119879119875

2 sdot 119879119875 + 119865119873 + 119865119875 (21)
















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia















References









































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia
































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


density-based statistical clustering: enabling sidefire...

Documents