您的当前位置：首页 A Traffic Object Detection System for Road Traffic Measurement and Management Abstract

A Traffic Object Detection System for Road Traffic Measurement and Management Abstract

来源：乌哈旅游

ATrafﬁcObjectDetectionSystemforRoadTrafﬁc

MeasurementandManagement

CarstenDalaff

InstituteforTransportResearch,

GermanAerospaceCenterDLR,Berlin,Germany

Carsten.Dalaff@dlr.de

RalfReulke

InstituteforPhotogrammetry,StuttgartUniversity,Germany

Ralf.Reulke@ifp.uni-stuttgart.de

AxelKroen

ISSPConsult,Stuttgart,Germanykroen@stgt.ssp-consult.de

ThomasKahl

ASISGmbH,Berlin,GermanyThomas.Kahl@asis-it.de

MartinRuhe,AdrianSchischmanow,GeraldSchlotzhauer,WolframTuchscheerer

GermanAerospaceCenterDLR,Berlin,Germany

ﬁrstname.lastname@dlr.de

Abstract

OISisanewOpticalInformationSystemforroadtrafﬁcobservationandmanagement.Thecompletesystemarchitecturefromthesensorforautomatictrafﬁcdetectionuptothetrafﬁclightmanagementforawideareaisdesignedundertherequirementsofanintelligenttransportationsystem.Particularfeaturesofthissystemarethevisionsensorswithintegratedcomputationalandreal-timecapabilities,real-timealgorithmsforimageprocessingandanewapproachfordynamictrafﬁclightmanagementforasingleintersectionaswellasforawidearea.Thedevelopedreal-timealgorithmsforimageprocessingextracttrafﬁcdataevenatnightandunderbadweatherconditions.Thisapproachopenstheopportunitytoidentifyandspecifyeachtrafﬁcobject,itslocation,itsspeedandotherimportantobjectinformation.Furthermore,thealgorithmsareabletoidentifyaccidents,andnon-motorizedtrafﬁclikepedestriansandbicyclists.Combiningallthesesingleinformationthesystemcreatesnewderivateandconsolidatedinformation.Thisleadstoanewandmorecompleteviewonthetrafﬁcsituationofanintersection.Onlybythisadynamicandnearreal-timetrafﬁclightmanagementispossible.Tooptimizeawideareatrafﬁcmanagementitisnecessarytoimprovethemodellingandforecastingoftrafﬁcﬂow.ThereforetheinformationofthecurrentOrigin-Destination(OD)ﬂowisessentially.TakingthisintoaccountOISalsoincludesanapproachforanonymousvehiclerecognition.Thisapproachisbasedonsingleobjectcharacteristics,orderofobjectsandforecastinformation,whichwillbeobtainedfromintersectiontointersection.

Keywords:Trafﬁcobservation,trafﬁccontrol,sensornetwork,sensorfusion1Introduction

Trafﬁcobservation,controlandreal-timemanagementisoneofthemajorcomponentswithinfutureintelligenttransportationsystems(ITS).OnecentralpostulationoftheEuropeanGovernmentistheincreaseofroadsafety,sothenumberofkilledpeopleshouldbehalvedtilltheyear2010.Therearenearly41.900roadcasual-tiesandmorethan1.7millionseriouslyinjuredpersonseachyearintheEuropeanUnion(EU).Thiscausesabout45billionEurodirectandapproximately160bil-lionEuroexternalcostsperyear.Daily4.000kmoftrafﬁccongestionsstressonlytheEuropeanhighways.Thismeans10%ofthecompleteEuropeanhighwaysystem.Theeconomicaldamageistremendous,buttillnowthereisnocommonapproachtocalculatedtherealamount.Sotheofﬁcialnumbersdiffer.ThenecessaryinvestmentsfortheEuropeantransportationﬁeldwillreachmorethan10%oftheEUgrossna-tionalproduct(GNP).Theneededﬁnancialcapabilitiesfortransportinfrastructureoftheaccedingcountries(e.g.Poland,Estonia)willincreasethisexpenseenor-

78Image and Vision Computing NZ

mously.Torealizejustthepriorityprojectsinthesecountries,theEUhastobespend91billionEuroupto2015.Takingthisintoaccountitseemseffectivetousepartsofthismoneyforinnovativeapproachesfortrafﬁcmanagementinfutureintelligenttransporta-tionsystems.Havingsuchanintelligenttransportationsystemanincreasedroadsafetycanberealized.Sotheeconomicaldamagecanbereduced.Oneappropri-ateapproachcouldbetheuseofthementionedtrafﬁcobjectdetectionsystemforroadtrafﬁcmeasurementandmanagement.Thissystemsisdifferenttostate-of-the-arttrafﬁcmeasurementequipment,e.g.inductionloops,whichdoesnotsufﬁceanymorethegrowingde-mandoftransportresearchandtrafﬁccontrol.TheprojectOIS[1]usesopticalandinformationalen-ablingtechnologiesforanautomatictrafﬁcdatagen-erationwithanimageprocessingapproach.Itsmainpurposeistoacquireandevaluateautonomouslytrafﬁcimagesequencesfromroadsidecameras.Trafﬁcpa-rameterswillbeobtainedfromextractedandcharac-terizedobjectsofthisimagesequences.Tomeettheserequirements,numerousimageprocessingalgorithmshavebeendevelopedsincemorethen20years(e.g.specialissue[2]),withsimpleweb-camerasandmorecomplexsystems(e.g.[3]).

Trafﬁcsceneinformationcanbeusedtooptimizetrafﬁcﬂowonintersectionsduringbusyperiods,identifystalledvehiclesandaccidents,andisabletoidentifynon-motorizedtrafﬁclikepedestriansandbicyclists.AdditionalcontributionscanbeobtainedforthedeterminationoftheOrigin-Destination(OD)matrix.TheODmatrixcontainstheinformationwhereandwhenthetrafﬁcparticipantsstartandendtheirtripandwhichroutetheyhavechosen.ODmatrixisonebasicelementforanoptimizedmodellingandforecastoftrafﬁcﬂow.Theestimatedtrafﬁcﬂowisnecessaryforadynamicwideareatrafﬁccontrol,managementandtravelguidance.

Furthermore,therecentadvancesincomputationalhardwarecanprovidehighcomputationalpowerwithfastnetworkingfacilitiesatanaffordableprice.Theavailabilityofspeciﬁcsolutionsinthelow-costgeneral-purposerangeallowsspecialimageprocessingandavoidssomebasicbottlenecks.Acoupleoftrafﬁcdatameasurementsystemsalreadyexist.Bestknownistheinductionloop.Inductionloopsareembeddedinthepavement.Theyareabletomeasurethepresentofavehicle,itsspeedandroughclassiﬁcation.Theseareonlylocalinformationbutforawideareatrafﬁcmanagementabigcoverageoftheareaisneeded.Anotherapproachisbasedontheideathatmovingve-hiclestransmitinformationabouttherepositionandve-locityviamobilecommunication,e.g.GSMtoatrafﬁcmanagementcenter.ThesedataarecalledFloatingCarData(FCD).Togetanoverviewofthetrafﬁcsituationofacompletecityatanytimeahugenumberofve-

Palmerston North, November 2003hicleshastobeequippedwithhardwareandmobilecommunicationunits.Togetreliabledataeveryoneminutethecurrentpositionandvelocityofthevehiclesisneeded.Thiscausesenormouscostsforthemobilecommunication.TheFCDapproachprovidesspatialtrafﬁcinformation,butthespatialandtimeresolutiondoesntﬁttherequirementofatrafﬁcsignalcontrol.OISasanewandinnovativetrafﬁcobservationsystemthatopenstheopportunitytodeliverallnecessaryinputforalocaltrafﬁcsignalcontrolaswellasforadynamicwideareatrafﬁcmanagementsystem.NextchallengeistheimplementationofOISinawideareacity.

2SystemRequirements

Amodernsystemfortrafﬁccontrolandreal-timeman-agementhastomeetthefollowingrequirements:•Reliabilityunderallilluminationandweatherconditions

•Workingperiodnon-stop,24hours7daysaweek•completeoverviewovertheintersectionfromatleast20minfrontto20mbehind•workinginreal-time

•real-time(everyhalfsecondacompletedatasetofthetrafﬁcsituation)onanintersection

Thisrequirementsshouldbetakenintoaccountforthedesignofallpartsofthesystem,whichcoversallpro-ceduresandprocessesfromimagedataacquisition,im-ageprocessingandtrafﬁcdataretrievaluptotrafﬁccontrol.Forrealizinganoperatingsystemfor24hoursandunderdifferentweatherconditionsinfraredcam-erasshouldbeused.Algorithmshadtobedevelopedforaspecialcameraarrangement(withreal-timede-mands)forvehicledetectionandderivingrelevantpa-rametersfortrafﬁcdescriptionandcontrol.

3SystemOverview

Togetacompleteoverviewovertheintersectionfromatleast20minfrontto20mbehinditisnecessarytohavemorethanonecamera.Thenumberofrequiredcamerasdependsontheintersectiongeometryandin-stallationpossibilitiesforthecamerasonhousewallsorlampposts.Thetimesynchronousimagedataacqui-sitionfromoneobservationpointwithdifferentcam-eraswillbedoneinaso-calledcameranode.ThecameranodeispartoftheOIS-philosophyandconsistsofdifferentsensorsasacombinationofVISandIR-cameras.Toﬁtthereal-timeprocessingrequirement,timeconsumingimageprocessingpartswillbeimple-mentedinaspecialrealtime-unit.Thisapproachisinimplementationatthemomentfortwointersectionsin

Berlin,Germanyequippedwithcamera-systems.Thearchitectureofthecompletesystemisshownintheﬁrstﬁgure.

Figure1:Systemdesignforthetestbed.Asshowninﬁg.1thecameranodesarepartofahier-archy,whicharelinkedviaInternetorWirelessLANwiththecomputersystemsontheintersectionandthemanagementcenter.Duetothelimiteddatatransferrateandpossiblefailureofasystemonanintersectionthecameranodeworksindependently.Imageprocess-ingisdecentralizedandwillbedoneinthecameranode,sothatonlyobjectsandobjectfeatureswillbetransmitted.Togetherwitherrorinformationtheobjectdatawillbecollectedonthenextlevelinthehierar-chy.Forsynchronizationpurposestimesignalscanbeincorporatedoveranetworkorfromindependentlyreceivedsignals(likeGPS).Thecameranodeispartofahierarchystartingwithcameranodes,junctions,sub-regions,etc.Thenextso-calledjunctionleveluniﬁesallcamerasobservingthesamejunction.Startingwiththeinformationfromthecameralevel,relevantdatasetsarefusedinordertodeterminethesameobjectsindifferentimages.Afterthat,theseobjectsaretrackedaslongastheyarevisible.Sotrafﬁcﬂowparameters(e.g.velocity,trafﬁcjam,cartracks)canberetrieved.Levelnumber3,theso-calledregionlevel,usestheextractedtrafﬁcﬂowparametersoutofthelowerlevelandfeedsthisdataintotrafﬁcmodelsinordertocontrolthetrafﬁc(e.g.switchingthetrafﬁclights).Additionallevelscanbeinserted.Fortestapplicationsthecam-eranodewillgenerateacompresseddatastream.Fortypicalworkingmodeonlyobjectfeaturesinanimageshouldbetransmittedandcollectedinacomputerofthehierarchy.Generally,opticalsensorsystemsinthevisibleandnear-infraredrangeoftheelectromagneticspectrumhavereachedaveryhighqualitystandard,whichmeettherequirementsevenforhigh-levelscien-tiﬁcandcommercialtasks,aboveallconcerningthera-diometricandgeometricresolutionanddatarate.Oth-erwisesensorsworkinginthethermalinfraredrange(TIR)arestillaresearchtopicfortrafﬁcapplications.Technologydevelopmentforthenextfewyearswillnotbefocusedonhigherresolutionsorfasterread-outs,becauseformostapplicationstheperformanceofthesesensorsissufﬁcient.Theemphasiswillbeputonsmart,intelligentsensorsystemswithdifferentmeasurement

80parameters(e.g.differentresolutions,differentspec-tralsensitivities)whichareconnectedwithinanetworksimilartotheinternetandbeingabletoconverttheincomingphysicalsignalsnotonlytodigitaldatabuttoprocessthemtouserneededinformation.Therefore,imagefusionaswellasfastandreliablealgorithmsareneededpreferablynearthesensoritself.Real-timeprocessingandprogrammablecircuitswillplayanim-portantrole.

4HardwareConcept

Thehardwareconceptisorientedontherequirements,asdescribedabove.Thesystemhastooperate24hoursand7daysaweekandtoobservethewholejunctionandatleast20mofallrelatedstreets.Toimprovetheopportunitiesforimageacquisitionmorethanonecamerasystemforoneobservationstandpointshouldbeinuse.Suchacameranodeﬁtstheﬁrstrequirementwithahigh(spatial)resolutioncameraandalowresolutionthermalinfraredcamera.Alsostereoanddistancemeasurementsarepossiblewithtwoidenticalcameras.Thecameranodeisabletoacquiredatafromuptofourcamerasinasynchronizedmode.Thecameranodehasreal-timedataprocessingcapabilitiesandallowssynchronouscaptureofGPSandINS(inertialnavigationsystem)data,whichshouldbeexternalmounted.Duetolimitationsofcameraobservingpositionsandpossibleocclusions(e.g.frombuildings,carsandotherdisturbingobjects)forintersectionobservationmostlymorethanonestandpointforacameranode(camera-system)isnecessary.Thereforecommunicationordatatransmissionbetweencameranodesandthecomputerinthenexthigherhierarchybecomescritical.Toreducedatavolumeinthenetworktherealtimedataprocessingcapabilityofthecamera-nodeisusedtospeedupimageprocessingandtotransmitonlyobjectdata.Forthereal-timeunitahardwareimplementationwaschosen.Largefreeprogrammablelogicgatearrays(FPGA)areavailablenow.Aprogramminglanguageisavailable(VHDL)anddifferentimageprocessingalgorithmsareimplemented.

5DataProcessing

Theprocessingofdatainthesensorwebfollowssuc-cessivesteps.Intheﬁrststepimagedataaregeneratedandpre-processed.Afterthattheobjectsareextractedoutoftheimages.Inthelaststepalltheobjectfeaturesfromdifferentcameranodeswillbecollected,uniﬁedandprocessedintotrafﬁcinformation.Thedatapro-cessingisoptimizedtothelogicaldesignofthesystem.Thereareoperationsforthesystemconﬁguration,fortheoperationalworkingforthedataminingandvisu-alization.Resultsarenewdataandinformation.Acomplexdatamanagementsystemregulatestheaccess,thetransmission,handlingandtheanalysisofthedata.

Image and Vision Computing NZ

5.1ImageProcessing

Theessentialprocessingstepsaretheelimina-tionofnoiseandsystematicerrors,compres-sion/decompression,higherlevelimageprocessingandspatial(orgeo-)-andtime-referencing.Thelastpointisnecessarytodeterminespaceandtimecoordinatesofobservedobjectsasanessentialfeaturefordatafusion.Mostoftheworkonvehicledetectionorrecognitionwasdoneongroundimages,mainlyaspre-processingbeforetrackingforsurveillanceortrafﬁcapplications[4].Differentapproaches,e.g.ﬁndingedges[5]ordeformablemodelforvehicle[?]canbefound.Generally,thereareproblemsinimageprocessingwithcarocclusion[6]andshadows[7].Besidestationaryimageacquisitionfromgroundalsofrommovingplatformsandarialimages[8]areused.Imageprocessingapproacheswillalsobeusedfordetectinglanesandobstaclesbyfusinginformation[9].ImageprocessingforOISwasdescribedindetailin[1].Themainprocessingisaclassicalobjectdetectionandidentiﬁcationtask.Followingmajorproblemshavetobesolved-Objectdiscriminationfromaspatialandtimevariablebackground(cross,street,buildings,etc.)-Removingdisturbingstructuresbetweenobjectandcamera,aswellasshadowregionsaroundtheobject-Identiﬁcationofcarsinarow,whichareoccludedbyotherTheﬁrstproblemcanbesolvedatleastintwodifferentways:

1.Workingwiththeimagesequences,whicharesubtractedfromtheimagebeforeor2.followingthetimechangingbackground.Fortheﬁrstapproachbackgroundcanbeeliminatedverysimple,butstalledvehiclesareinvisible.Theotherapproachisanon-goingupdateofthebackground.Thisneedsmuchmoreexpense,butallowsadetectionofmovingandalsostalledcars.Asaresultofthisprocedure,backgroundcanbesubtractedfromthecurrentimageandobjectscanbederivedasshowninﬁg.2.Additionalmorphologicaloperationremovesclutterandcloseobjectsstructures.Therightimageshowsthegreycodedobjectsafterlabelling.

Figure2:Objectdetectioninatrafﬁcscene(left:originalimage,rightprocessedimage.

Theothermajorproblemoccursafterdetectingtheob-ject.Todeterminesizeandshape,disturbingeffects

Palmerston North, November 2003likeshadowshavetoberemoved.Astraightforwardwayistheanalysisofgreyandcolorvalues,aswellastexturewithinthefoundobjectboundaries.

5.2Dataandobjectfusion

Toensureanoperatingsystem24hoursadayevenun-derbadweatherconditions,e.g.rain,fog,andatnightthefusionofavisibleandanIRsensorisapromisingapproach.Thefusionispossibleondifferentstagesofinformationprocessing:

-datalevel(e.g.imagedatafromdifferentcameras)-objectlevel(objectsextractedfromtheimagedata)

-informationlevel.

Imagematchingandregistrationasonepartofthedatafusionisaprocedurethatdeterminesthebestspatialﬁtbetweentwoormoreimagesacquiredatthesametimeanddepictingthesamescene,byidenticalordifferentsensors.Tofusethedifferentimagesand/orobjectdataasynchronizationoftimeisnecessary.Timesynchronizationcanberealizedbyinternalclocks(e.g.computer)orexternaltimeinformation(e.g.GPS).FortheapplicationofOISwemergeimagesoncameranodelevelandfusethepositioninformationandobjectfeaturesonjunctionorregionlevel.Theobjectinformationarefedinfromdifferentcameranodes.Bothproceduresshouldbeexplainedmoredetailed.Fusionondatalevel:Toﬁttherequirementfor24-hourobservations,typicalCCD-camerasfailbecauseoflimitedilluminationoftheobjects.Carheadlightsandrearlampsseemsnotbesufﬁcient.Toovercomethisproblemthecarself-radiation,whichhasamaximuminthethermalspectralregion(TIR),canbeused.Thereareabunchofdifferentdetectorssensitiveinthisspectralrange.Mostofthesensorsareexpensiveandneedsadditionalcooler.Recentdevelopmentsshow,thatbolometerarraysareacandidateforcheaperanduncooleddetectorarrays.Therefore,acameradevelopmentwasstarted,whichgivesfullaccesstothesensor,control,datacorrectionanddataﬂow.Firstexperimentsweredonewithacommercialsystem.Anexampleofthedataandthefusionofbothondatalevelisshowninﬁg.3.

Observationwasdoneonlateafternoon.TheleftimageisatypicalCCD-image.Contrastbecomessmaller,onlyreﬂectionsfromsunglittersoncarroofsarevisi-ble.Themiddleimagewastakenfromabolometersen-sor.Thewholeintersectionandthestreetarevisible.Therightimageisthefusionofboth.Forvisualiza-tiongreylevelimagewasputinthegreenchannelandtheTIRimageinthered.MergingthevisibleimageandtheIR-imageaafﬁnetransformationwasused.To

fusedatafromdifferentsourcesisobvious,butneedsspatialandtimesynchronization,becauseofdifferentimagingsystemandobservationconditions.Forthisexample,thesynchronizationtaskisbasedonmanualprocedures,likeﬁndingequivalentpointsinbothIRandVISimagesandthecalculationofthenecessarytransformation.Especiallytheautomaticspatialsyn-chronizationisaresearchtopic.Alltheseoperationsaredoneineachcameranode.AnotheradvantageofTIRimagesshowsthefollowingimagepair,whichwastakenfromthesameobservationpoint,butatdaytime.

Figure3:SensorfusionofVISanIRimages(I).

Figure4:SensorfusionofVISanIRimages(II).Fig.4showsanexampleoffusingRGBandTIRim-agesatdaytime.AftercoregistrationtheIR-imagetotheRGB-imageandapplyingafﬁnetransform,adirectcomparisonispossible.Indifferencetoﬁgure1theinfraredimagewasﬁtdirectlyintothegreylevelim-age(acolorseparationoftheRGB-image).TheRGBimagehasamuchmorehigherresolutionthantheTIR-image(720x576pixel).Spatialandtruecolorobjectdatacanbederivedfromthisimage.TheTIR-imagehasasmallerresolution(320x240pixel).Thecom-binationofbothimagesgivesamorecomplexresult,becauseofthethermalfeatures,whichcanbeobservedontheenginebonnetandasreﬂectedradiationfromunderthecar.Thesearealsonewfeaturesforobjectdetectionanddescriptionwithintheimage-processingtask.ObjectDataFusion:Occludedregionscanonlybeanalyzedwithadditionalviewsontheseobjects.Be-causecameranodesalwaystransmitobjectinforma-tion,adatafusionprocessonobjectlevelisneces-sary.Theobjectlistfromdifferentcameranodeshastobeanalyzedanduniﬁed.Objectfeatureslikeposition,size,andshapevaryfromeachviewangle.Thereforethedifferentobjectfeaturesfromdifferentperspectiveshavetobecompared.Theresultofthisoperationisonetrafﬁcobjectwithanexactposition,sizeandshape,etc.Theassumptionsforthisoperationaretimeandspatialsynchronizedimagedata.Thesequentialprocessingofthislistallowsthederivationofmoredetailedinfor-mation,e.g.trackofthetrafﬁcparticipants,removingerroneousobjects,etc.

825.3DataAcquisitionandGeoreferencing

Transformationfromimagetoworldcoordinatesisanessentialneedinordertocalculatemetricbasedtrafﬁcdata.Standardphotogrammetricproceduresforthetransformationofcoordinateswithinmonocularimagesareused.Basicassumptionsarewelldistributedandaccuratemeasuredgroundcontrolpoints(GCP)inobjectandimagespaceaswellasanexactcameracalibration.TheGCPsarecalculatedviaDGPSwithinWGS84andUTM-projection.Thecalculatedcameracalibrationparameters(interiorandexteriororientation)andimagecoordinatesareinputforthetransformationequations.DuetomonocularimageaccusationtheobjectorvehiclepositionshavetobeprojectedonaXY-planeinobjectspace.Thevehiclepositionsprojectedonthatplanedependonthecameradistance,thecameradeclinationandthepositionofthepointwithinthevehiclerepresentingitsposition.Oncethecameracalibrationissetthevehiclepositionscanbetransformedtoworldcoordinateswithinimagesequencesuntilthecamerapositionchanges.Additionally,intersectiongeometryforexamplelanesetc...canbetransformedfromimageintoobjectspace.

5.4CalculationofTrafﬁcCharacteristics

Thefollowingﬁgure5showsthegeneraltrafﬁcchar-acteristicscalculationprocess.

Figure5:Generaltrafﬁccharacteristicscalculationprocess.

Theimageprocessing(notshownabove)cyclicallydeliverstheparametersofallidentiﬁedtrafﬁcobjectssuchascars,trucks,cyclistsandpedestriansintrafﬁcobjectlists(TOLists),containingtype,size,speed,direction,geographiclocationetc.oftheobjectsforacertainsampletime.Astorageprocedure(TOstorage)writesthisrawtrafﬁcdataintoadatabaseforfurtherprocessing.Atrackingprocedure(TOTracking)markstrafﬁcobjectsappearinginconsecutivetimesamplesbyanuniqueobjectidentiﬁcation.Thus,trafﬁcobjectscanbepursuedthroughouttheobservedtrafﬁcarea.ThetrackedtrafﬁcobjectsandtheirparametersarestoredintheTrackedTODataareaofthedata

Image and Vision Computing NZ

base.Usingthisdata,atheTrackingCharacteristicsCalculationmodulecomputestrafﬁccharacteristics,e.g.trafﬁcdensityorﬂowrates.Tocomputethetrafﬁcdensitythenumberofmotorvehiclesonacertainroadsegmentisnecessary.Usingtheirgeographicalcoordinatesanddirectionallmotorvehiclesmovingalongaroadsegmentareselectedformthedatabase.Bysimplecountingthenumberofthisvehiclesandscalingtheirnumbertoaonekilometersegmentthetrafﬁcdensitycanbeobtained.Flowratescanbedeterminedbycountingthenumberofcarscrossingadeﬁnedtraversesection.Basedonthetrafﬁcobjectparametersetoveranumberofsampletimesandusingthetrackinginformation,thenumberofvehiclescrossingthesectioniscountedandatrafﬁcﬂowmeasureinvehiclesperhourorsocanbeobtained.

6AnNetworkAdaptiveControl

andHighDynamic

Thevideobasedtrafﬁcsensordevelopedinthisprojectcreatespossibilitiesfornewconceptsoftrafﬁccontrolforintersectionsandwideareanetworks.Thisapproachalsoincludesthedevelopmentofanewdynamicandadaptivetrafﬁccontrolmodelsfortrafﬁclights.State-of-the-artintrafﬁcobservationistheinductionloop.Inductionloopsareembeddedinthepavementandregisterabout1·2m.Thiskindofsensorisabletomeasurethepresentofavehicle,itsspeedandroughclassiﬁcation.Inanextprocessingstepitispossibletocalculatetimeintervalsbetweenvehicles.Thisdataisneededtocontroltrafﬁclightsignalsonintersections.Forarealdynamicanddemandbasedtrafﬁcsignallightcontrol,youwouldneedthedataofnumerousinductionloopsononesingleintersection.Thisisneitherefﬁcientnorrealizable.OISsensorweboffersanewkindoftrafﬁcdata/information.Itisbasedonsocalledtrafﬁc-actuatedsignals.Thismeans,thesystemisdetectinginformationaboutthereal-timetrafﬁcsituationonanintersectionautomatically.Forexamplethelengthofthequeueforalldifferentlanes,thetrafﬁcﬂowattheintersection,thedensityoftrafﬁcandthecurrentvelocityofeachvehicle.TheOISsensorwebautomaticallyprocessesdata-position,speedvectorsofeachvehicle,queuelengthaswellasotherrelevantfeatures.Thisleadstoacompleteandreal-timeoverviewatleast20mbeforeandbehindanintersection.Basedonthisnewqualityofdata,newapproachesaredevelopedtocontroltrafﬁclightsignalsondynamicdemand.Atthemomentmostoftrafﬁclightsarecontrolledbytwodifferentways:1.controlbyﬁx-timesignals2.controlbysocalledactuatedsignalsFix-timesignalsmeans:greenandredtimeisﬁxedoverthetimeandindependentoftheactualsituationontheintersection.Actuatedsignalsmeans:anumberofﬁx-timesignalsareusedfordifferentdemandsandsituations.Atthistimetherearenearly

Palmerston North, November 2003nosensorwebsormeasurementsystemsavailable,thatareabletomeasureeachobjectonaintersectionandthatprovidesallnecessaryspatialinformationforarealdynamictrafﬁccontrolsystem.OISisasystemthatisabletoacquirefeatureslikesize,shapeandotherobjectfeaturestoclassifyandidentifytrafﬁcobjects.Summarizingthis,OISgivesacompletedatainformation(overview)overawholeintersection.TheprojectisfundedbytheGermangovernment(FederalMinistryofEducationandResearch,registrationnumber:03WKJ02B).

References

[1]R.Reulke,A.Bœrner,H.Hetzheim,A.Schis-chmanow,andH.Venus.Asensorwebforroad-trafﬁcobservation.InImageandVisionComput-ing,NewZealand2002,pages293–298,2002.[2]A.BroggiandE.D.Dickmanns.Applicationsof

computervisiontointelligentvehicles.ImageandVisionComputing,18(5):365–366,April2000.[3]F.Pedersini,A.Sarti,andS.Tubaro.Multi-cameraparametertracking.IEEProceedings-Vision,ImageandSignalProcessing,148(1):70–77,February2001.[4]M.Betke,E.Haritaoglu,andL.S.Davis.Real-timemultiplevehicledetectionandtrackingfromamovingvehicle.MachineVisionandApplications,12(2):69–83,2000.[5]J.Canny.Acomputationalapproachtoedge

detection.PAMI,8:679–698,1986.[6]I.Ikeda,S.Ohnaka,andM.Mizoguchi.Traf-ﬁcmeasurementwitharoadsidevisionsystem-individualtrackingofoverlappedvehicles.InProceedingsofthe13thInternationalConferenceonPatternRecognition,pages859–864,1996.[7]G.S.K.Fung,NelsonH.C.Yung,GranthamK.H.

Pang,andAndrewH.S.Lai.Effectivemovingcastshadowdetectionformonocularcolrtrafﬁcimagesequences.OpticalEngineering,41:1425–1440,June2002.[8]R.Chellappa.Anintegratedsystemforsitemodel

supportedmonitoringoftransportationactivitiesinaerialimages.InProceedingsofthe1996ARPAImageUnderstandingWorkshop,pages275–304,1996.[9]M.BeauvaisandS.Lakshmanan.Clark:a

heterogeneoussensorfusionmethodforﬁndinglanesandobstacles.ImageandVisionComputing,18(5):397–4,April2000.

因篇幅问题不能全部显示，请点此查看更多更全内容

查看全文