PaperITS2015ProbabilisticModelsforSensorSimulationsfinalpdf智能交通世界大會(huì)ITS智慧城社區(qū)人工智能A

PaperlTS2015ProbabilisticModelsforSensorSimulationsfinal.pdf 智能交通世界大會(huì)ITS智慧城市社區(qū)人工智能AI物聯(lián)網(wǎng)IT報(bào)告課件教案22ndITSWorldCongress,Bordeaux,France,59October2019PapernumberITS-2627Probabilistic SensorSimulationsforValidatingDataFusionSystemsRobinSchubert1*,NormanMattern1,RobinvanderMade21.BASELABSGmbH,Ebertstr.10,09126Chemnitz,Germany,robin.schubert@baselabs.de2.TASSInternational,TheNetherlands AbstractWiththeincreasingdeploymentofadvaneeddriverassistaneesystemsandtheongoingdevelopmentofvehicleautomation,efficientwaysofvalidating suchsystemsarebecomingacrucialpartofthedevel-opmentprocess.Inparticular,simulationsareanincreasingly important addition tofieldtrialsastheyfacilitateanearlyandautomatedevaluation.Inthispaper,aprobabilisticmethodologyforsimulatingsensordatainthecontextofadvaneeddriverassistaneesystemsandautomatedvehiclesispresented.Theobjectiveofthisapproachistoincreasethesimulationslevelofrealismwhilemaintainingbothflexibilityandadaptabilityof simulation-basedvalidation strategies.Theproposedprobabilistic sensormodelsarecomparedtorealradardatainordertoevaluatethestatisticalcharacteristicsofbothdatasets.Withthepresentedapproach,itwillbepossibletoincreasethequalityoftheinitialevaluationresultsbasedonsimulateddata.Keywords: Sensorsimulation, MonteCarlo,ProbabilisticfilteringIntroductionInordertofurtherincreaseroadsafety andtrafficefficiency,advaneeddriver assistaneesystemsarecurrently beingwidelydeployed.Inaddition,different stakeholdersarecurrently investigating howanincreasinglevelofvehicleautomationcancontributetotheseobjectives[1].Asthesesystemsaredi-rectly intervening intothedriving process,theirdesignandimplementationishighlysafety-critical.Appropriateevaluationmethodologiesareacrucialpartofanydevelopmentprocessforsuchsystems.Duetothehighcomplexityoftrafficscenarios,fieldtrialsrequireatremendouseffort including driving millions ofkilometres.Thus,evaluationmethodologiesbasedonsimulationareincreasinglyappliedProbabilisticSensorSimulationsforValidatingDataFusionSystems2inparticular,fortheearlyphasesofevaluation.Themainbenefitsofsimulationsinclude thepossibilitytoautomatetests,toconductevaluationseveniftheplatform(e.g.sensors)arenotyetavailableandtoassesssafety-criticalsituations.Ontheotherhand,thesignificaneeofsimulation-basedevaluationsstronglydependsonthequalityofthesimulations,thatis,ontheprobabilitythatrealandsimulatedtrafficseenarioswouldtriggerasimilarbehaviourofthesystemundertest.Currently,twomainapproachesofsimulatingsensordataarebeingused:Groundtruthsensormodels:Thesemodelsdeliverthetrue,undisturbedsimulatedvaluesofthesimulatedquantities(e.g.,thepositionandvelocityofvehiclesorthecurvatureofaIane).Thenotionbehindthiskindofmodelsisthatasystemwhichfailsonidealizeddatawillcertainlynotfulfilitsrequirementsinrealisticscenarios.Physics-basedsensormodels:Thesemodelsattempttocovertheinternalbehaviourofthesensorandthephysicalmeasurementprinciple.Asanexample,manysimulationenvironmentsproviderenderedcameraimagesthataccount,amongothers,forlightingandweatherconditions. Similarly,physicalradarsensorsexistthatcalculatethepropagationofelectromagneticwavesinthetrafficsceneandthedetectioncharacteristics(e.g.,theantennapatterns) orthesensor.Whileeachoftheseapproachesisjustifiedforcertainusecases,bothlevelsofmodelling haveparticulardrawbacks.Thedisadvantageofgroundtruthmodelsisratherobvious,astheycompletelyneglect sen-sordisturbanceswhichdeterioratesthesignificanee oftheevaluationresultsobtainedwithsuchmodels.Thoughphysicalmodelsappeartoovercomethislimitationbymaximizingtherealismofthesimulateddata,theirdrawbacksareratheraveryhighcomputationalcomplexityandevenmore importantaratherlimitedpossibilitytoadaptthesimulationtodifferent sensortypes.Infact,exchanging,e.g.,aDopplerradarbyafrequencymodulatedcontinuouswave(FMCW)radarimpliestodevelopacom-pletelynewphysicalsensormodel.Table1ComparisonofdifferentabstractionlayersofsensormodelsforsimulationCriteriaGroundTruthModelsPhysicalModelsProbabilisticModelsErrorCharacteristicsidealizedrealisticrealisticstatisticsComputationalComplexitylowveryhighLowAdaptabilitytospecific sensorsn/averyhighlowProbabilisticSensorSimulationsforValidatingDataFusionSystems3Figure1GeneralstructureoftheprobabilisticsensormodelapproachInthispaper,anintermediateabstractionlayerforsensorsimulations ispresentedwhichintegrates sen-sordisturbancesprobabilistically.Thus,theobjectiveistherepresenttheerrorstatisticsofrealsensordataratherthanthedatathemselves.Table1givesacomparisonofthisapproachandthetwoclassicalmodellinglayers.Thepaperdescribesthetechnicalapproachandpresentsfirstresultsthathavebeenobtainedbycomparingprobabilisticallysimulateddatatorealdatainatypicaltraffic scene. Technical approachandchallengesThe generalideaofthepresentedapproachthatisillustratedinfigure1appearsratherstraightforward:Theidealizedsensordatageneratedfromagroundtruthsensormodelaresuperimposedbyanerrorsignalusingarandomgenerator.Inpractice,thiscanbedoneusingaMonteCarloapproach(forinstanee, rejection sampling[2]).Thisapproachcanbeappliedtodifferenttypesofsensorerrors,includingSensornoiseforeachmeasuredguanLily,l?alsenegativedetectionSjl;alsepositivedetecLions,Timingenvors(deterministic/probabilistic sensorlatencies)Themajorchallengeistoselectanappropriateprobabilisticdensityfunction(PDF)tosamplefrom.ThisPDFneedstorepresenttherealcharacteristicsofthesensorwhilestillfacilitatingadaptability.Thisadaptabilityshallnotonlycoverdifferent

sensors,butalsodifferent environments,weatherconditions,etc.Thistrade-offisachievedbydefiningaparticulartypeofPDFforeacherrortype(e.g.aPoissondistributionfordetectionerrororaRayleigh distribution forradardetections).However,theparametersofthesePDFs(e.g.,theclutterdensityforaPoissondistribution)canstillbesetaccordingtothesensortoberepresentedorthecurrentscenario.IdealizedSensorDatafromaccordingtothesensortoberepresentedorthecurrentscenario.IdealizedSensorDatafromSimulationProbabilisticSensorModelsSimulationSimulationProbabilisticSensorModelsSimulationEnvironmentSensor DatawithrealisticerrorcharacteristicsProbabilisticSensor SimulationsforEnvironmentSensor DatawithrealisticerrorcharacteristicsProbabilisticSensor SimulationsforValidating DataFusionSystems4CaseStudyInordertocomparetheprobabilisticallysimulatedsensordatawithrealdata,thefollowingevaluationmethodologyhasbeenapplied:DatafromvarioussensorshavebeenrecordedusingthedatahandlingframeworkBASELABSConnect[3].Thedataincludescameraimagesanddetectionsofa77GHzFMCWradar,Fromtherecordeddata,asimulationsscenariohasbeenderivedusingthesimulationsoftwarePresScan[4].Vehiclesinfrontoftheegovehiclehavebeensimulatedusingagroundtruthpositionandvelocitysensor(cp.figures2).Figure2Comparisonofrealandsimulatedtraffic scenariousedfortheevaluation. Figure3:IdealizedandmodifiedradarmeasurementsProbabilisticSensorSimulationsforValidatingDatafusionSystems5Usingtheapproachpresentedinthispaper,sensornoisehasbeenaddedtotherange,rangerate,andazimuthmeasurementsoftheradargroundtruthdata.Inaddition, detectionerrorsincludingfalsenegativesandfalseposiLives(c1ntter)hriveheenridded. Th^已mncharacteristicsoftheprobabilisticsensormodelshavebeencomparedtothestatisticsoftherealsradardata(includingthedetectionperformaneeandthemeasurementaccuracy)asshowninfigure3.Thecomparisonshowsthatthesimulateddisturbeddatarepresentsthesta-tisticalcharacteristicsofthetruedatareasonablywellwhichdoesnotappearsurprising,astheparametersoftherandomgeneratorhavebeenderivedfromtheseverymeasurementsbefore.Thisexemplaryevaluationshowsthatitiscomparablyeasytogenerate simulateddisturbedsensordataifthestatisticalpropertiesofthesensorundertestarewellknown.ResultsIn additiontothequalitativeevaluationdescribedintheprevioussection,aquantitativevalidationhas beenconducted.Theobjectivewastoensurethatthestatisticalpropertiesthataresupposedtobemod-elledcanbeindeedfoundinthesimulatedsensordata.Inthefollowing,theresultsforthedetectionerrorsarepresented:Forfalsenegatives,theuseroftheprobabilisticsensormodelmaydefinethedetection probabilityofthesimulatedsensor.Fromallsimulateddetections,asubsetischosesprobabilisticallythatissimulatedasnotdetectedand,thus,isnotdeliveredtotheoutputofthesimulationmodel.Infigure4,thecumulatedratiobetweenthedetectedobjectsandtheexistingobjectsisillustrated.Forthisexperiment,aparameterof=0.7hasbeenused.Itcanbeobservedthatwhileatthebeginningofthesimulation,theresultingratioisratherdynamic,itisconvergingagainst70%duringthesimula-tion.Thevalidationofthefalsepositivedetectionsrequiresabitmoreofexplanation:Themainparameterofthesimulationforthiseffectisthenumberoffalsepositivedetectionswithinthefieldofview.Thisparameterisnotprobabilisticonthecontrary,itisadeterministicvalue(whichmeansthatifthevalueissetto2,exactly2falsepositivedetectionsaresimulatedateachtimestep.However,thepositionsofthefalsepositive detectionsarechosenprobabilistically.ConsidertheexampleofanACCshowninfigure5:Theegovehicleisadjustingitsspeedaccordingtothedistaneeandthevelocityofthetargetvehicleinfrontofhim.Afalsepositivedeteetionontheneighbourlaneandinfrontofthetargetvehiclewillnotchangethebehaviourofthesystem.However,afalsepositivedetectionbetweentheegoandthehostvehiclewillhaveaneffect.Thus,theareaoftheegolanebetweenbothvehiclescanbeconsideredanareaofinterest fortheACCwithrespecttofalsepositivedetections.ProbabilisticSensorSimulationsforValidatingDataFusionSystems6Thequestionishowmanyfalsepositivedetectionswilloccurwithinthisareaofinterest.DatafusionsystemstypicallyassumethatthenumberoffalsepositivedetectionsisfollowingaPoissondistribution, whosedensitycanbecalculatedbymultiplyingthenumberoffalsepositivedetectionswiththeratiobetweentheareaofinterestandtheareaofthefieldofview.Figure6showsboththetheoreticalPoissondistributionforthegivenscenarioaswellastheempiricalvalues.Itcanbeseenthatthesimulationfitswelltothetheoreticalassumptions.Thevalidation showsthatthesimulatedmeasurementsbehaveaccordingtotheassumptionstypicaldatafusionsystemshave(thatis,Gaussiannoise,adefineddetectionprobabilityandafalsepositivedensitythatfollowsaPoissondistribution).Thus,thesimulationcanbeconvenientlyusedtotestandvalidatedatafusionsystemsanddeterminetheirbehaviourunderthecondition thattheirassumptionshold.Fu-tureworkwillalsoincludethesimulationofeffectsthatviolatessuchassumptions. ProbabilisticSensorSimulations forValidating DataFusionSystems7Figure4:Simulatedfalsenegativedetections.Inthetopdiagram,thetimestepsfrom0to500are