Multisource-Traffic-Data-Fusion-with-Entropy-Based

MultisourcetrafficdatafusionwithentropybasedmethodSunZhanquan,GuoMu,LiuWei,FengJinqiaoShandongComputerScienceCenter,Jinan,250014,[e-mail]sityofTechnology,Dalian,116400,ChinaAbstract—ItisacrucialpartforATMStoaccuratelyidentifyandforecasttrafficstatefromreal-timetrafficdata.Toimprovetheidentificationrateoftrafficstate,multisourceinformationshouldbeused.Themultisourceinformationfusionmethodisimportant.Informationfusionisdividedintothreelevels,i.e.datalevel,featurelevel,anddecisionlevel.Intrafficcongestionidentification,manymeanscollectedtrafficdatasourcecanbeused,suchasinduceloopvehicledetector,videodetector,GPSfloatingcarandsoon.Thetrafficstatecanbeidentifiedaccordingtoeachdatasource.Forimprovingtheidentificationrate,wedevelopadecisionlevelmultisourcefusionmethod.Inthemethod,Bayesianinferenceisusedtoobtainthetrafficstateinprobabilitystyleaccordingtoeachdatasource,andentropybasedweightedmethodisusedtofusetheresultindecisionleveltoimprovetheidentificationrate.Theentropybasedfusionmodelandalgorithmisintroducedandpresentedinthispaper.FielddatacollectedthroughloopvehicledetectorandGPSfloatingcarareanalyzedwiththeproposedmethod.Keywords-Intelligenttransportationsystems;datafusion;trafficdata;entropyI.INTRODUCTIONRecognitionoftrafficcongestionisimportantforvehiclenavigation,trafficcontrolandmanagement,aswellasformanyotherintelligenttransportsystemsintraffic.Ithasbecomeoneofthemajorissuesinmostcountriesandhasbeenwidelystudied[1-2].Currently,trafficcongestionidentificationcanbeperformedthroughthreemeans,i.e.videodetection,patroldetection,andtrafficflowparameterbaseddetection.Trafficflowparameterbaseddetectionmethodhasbeenwidelyacceptedbecausethatitisunaffectedbyweathercircumstanceandcanbeperformedautomatically.Manyidentificationmethodsbasedontrafficflowparametershavebeenstudied,suchasCalifornia,McMaster,neuralnetwork,Bayesianinference,fuzzylogicmethodsandsoon[3-9].Eachmethodhasitslimitation.Forimprovingtrafficcongestionidentificationrate,multisourcedatafusionisthebestchoice.Sometrafficdatafusionmethodshavebeenstudied[10-11].Theobjectiveofmultisourcedatafusionistominimizetheimpactofuncertaintiesandgettheinformationoutofsources.Generallyspeaking,amultisourcedatafusionapproachinvolvesthreecomponents:informationrepresentation,uncertaintydescriptionandoptimization.Afusionprocesscanbecarriedoutatvariouslevelssuchasdata,feature,anddecisionlevels.Datalevelfusioncombinesrowdatareceivedfromdifferentsensorstoprovideaconsensusdata.Featurelevelfusionconsiderscombiningthefeaturesextractedfromrawdata.Decisionlevelfusionconsiderscombiningthedecisionobtainedaccordingtoeachdatasource.Inthispaper,wedevelopanefficientalgorithmforsolvingthemultisourcedatafusionproblemindecisionlevel.TwocurrentlyimportantcoexistingsensortechnologiesareforinstancetheinductiveloopdetectorsandGPSfloatingcar[12].Loopdetectorsmeasurethetrafficprocesstemporally.Itprovidesadataqualityinaccordancewiththeirphysicalfunctionalprincipleandinaccordancewiththeinfluencesoftheaffectingsurroundingenvironment.Relatively,trafficflowdatacollectedthroughloopdetectorismorereliablebecauseitisunaffectedbyweather.Itcanprovidespeed,timeoccupancy,andtrafficvolumeetc.parameters.Itisthecommonlyusedtrafficdatacollectionequipment.Thedisadvantageofloopvehicledetectoristhatthecoverageislow.Floatingcarcanprovidereal-timenetworkinformationbytransmittingtheirspeed,positionanddirection.Usingthisdata,trafficcongestioncanbeidentified,traveltimescalculatedandtrafficreportsgenerated.Becausethefloatingcardistributedontheurbanroadnetworkrandomly,thecoverageisveryhigh.Floatingcardataisthemaindatasourceoftrafficcongestionidentificationcurrently.Intheproposedmultisourcedatafusionmethod,Bayesianinferencemethodisusedtoidentifytrafficstateaccordingtoeachtrafficdatasource.Mutualinformationbetweentheidentificationresultandrealtrafficstateisusedtocalculatetheweightofeachdatasource.Mutualinformationcanmeasurearbitrarycorrelationofvariables[13].Theweightbasedonmutualinformationcandisplaytherealcorrelationofeachdatasource.ThetrafficstateisidentifiedwithBayesianinferencemethodaccordingtofusedprobabilitydistribution.TheefficiencyofthemethodisillustratedthroughanalyzingJinanurbantrafficdata.II.DATAPREPROCESSINGTrafficcongestioniscommonlyidentifiedaccordingtotrafficflowdatacollectedbydifferentkindofdetectors,suchasloopvehicledetector,ultrasonicradardetector,andmicrowavedetectorsandsoon.Trafficflowdataisthebasisofallkindofintelligenttransportationanalysis.Thedataqualityhasgreateffectontheidentificationrateoftrafficcongestion.Forimprovingtheidentificationrate,wehavetopreprocessthecollectedtrafficdatatoguaranteethedataquality.Therecommonlyexistthreekindsoffalsedataamongcollectedtrafficdata,i.e.imprecisedata,errordata,andmissingdata.Imprecisedataiscausedbyenvironmentnoise.Errordataisusuallycausedbydetector’sfault.2009InternationalConferenceonArtificialIntelligenceandComputationalIntelligence978-0-7695-3816-7/09$26.00©2009IEEEDOI10.1109/AICI.2009.392506Missingdataisusuallycausedbydete