Multi-Target-Multi-Object-Tracking-Sensor-Fusion-o

2004IEEEIntelligentVehiclesSymposiumUniversityofParmaParma,ItalyJune14-17,2004Multi-TargetMulti-ObjectTracking,SensorFusionofRadarandInfraredRainerMobus,'*'Research&Technology,RICIAADaimlerChryslerAG,70546Stuttgart,Germanyemail{rainer.moebus,u1i.u.kolbe}@daimlerchrysler.comAbstract-Thispaperpresentsalgorithmsandtechniquesforsingle-sensortrackingandmulti-sensorfusionofinfraredandradardata.Multiplemodelfilteringanddataassocia-tiontechniquesarepresentedandresultsareshownforallpresentedalgorithms.I.INTRODUCTIONManydriverassistancesystemsheavilyrelyonexternalvehiclesensors.Oneproductthatisalreadyavailableonthemarketistheso-calledAdaptiveCruiseControl,ACC,thatmakesuseofa77GHzradarsensor.Themaximumac-celerationisconstrainedto12,themaximumdecelerationislimitedto-32.Aconsiderabledisadvantageofcurrent77GHzACCsystemsistheirlimitedangularsensorrange.Thelongitudinalrangeisverygoodandgoesupto150Fig.1.Sensorrangesof77GHzradarandinfraredlaser:meters,inbestcasesbeyond.Thelateralrange,however,islimitedtof3.4degrees.ForassistancesystemsbeyondACCthe77GHzradarsensor.mightnotbesufficient.ForimprovedACCfunctionality,lanechangeassistants,inner-cityassistingsystemsorcollisionwarning/collisionmitigationsystemsarathercompleterepresentationofthetrafficsceneaheadofthecarisneededandthereliabilityofthesensorinformationshouldbeashighaspossible.Anothersensorsuitableforobjectrecognitionistheinfraredsensor,scanninganareaofaboutf20degreesupto80meters.Figure1showstherangesofthe77GHzradarsensor'andtheinfraredscanner,theinfraredsensorscanningalargeareauncoveredbytheradar.Toexploittheadvan-tagesofeachtechnologywithoutbeingaffectedbytheirshortcomingsallsensorinformationcanbefusedinasensorfusionstep.Complementaryregionsadduptoacompleterangeaheadofthecar,inoverlappingregionsredundantinformationisobtainedincreasingthereliabilityoftheUliKolbe'*AutomaticControlLaboratoryETHZentrum-ETL,CH-8092Zurich,[e-mail]2showsa7'7GHzstandardACCsensorandtheinfraredsensorusedforthemulti-sensorfusioninthispaper.Toshowthemotivationandnecessityoftrackingandfusionalgorithmsfigure3showsabirdviewofallinfraredandradarmeasurementsobtainedduringa20secondtestdrive.Thegraphshowsthetrajectoryofthetestvehicleasasolidlineandtheobtainedsensormeasurementsaspoints.Forthesakeofreadablitythelateralpositionofthemeasurementsisexaggeratedbyafactorof20.Thesedatashowthatapowerfultrackingandpowerfuldataassociationandfusionalgorithmsareneededtorejectclutterandnoiseandtoestimatethedynamicsoftheprecedingvehiclesreliably.Fig.2.automotiveuse.Picturesofa77GHzradarsensorandaninfraredsensorfor11.OUTLINEOFTHISPAPERThispaperpresentsalgorithmsandtechniquesforthesingle-sensortrackingandthemulti-sensorfusion.Chapterthreeisdedicatedtosingle-sensortrackingandtoresultsofradartrackingandinfraredtrackingwiththeinteractingmultiplemodelapproach.Chapterfourcoverssensorfusionanddataassociationtechniques,especiallytheprobabilisticdataassociation.Thefifthchapterpresentsresultsofmulti-plemodelfilteringappliedtothesensorfusionofinfraredandradardata.Thelastchaptergivesanoutlook.111.SINGLESENSORTRACKINGAnoverviewoftheoperationsperformedinthesingle-sensortrackingisshowninfigure4.Differentsensorsgeneratesensorspecificmeasurementsortargets.Inthefirstoperation(gating,clustering)apre-selectionofthesetargetstakesplace.Newtargetsarecomparedtoexistingo-7ao3.~a31o-91o41$2o.ooo2004IEEE732150-100-500so100150200lateralposition[m]Fig.3.vehicleaswellasallinfraredandradarmeasurententsareshown.Birdviewofa20secondtestdrive.ThetrajectoryofthetestFig.4.trackingstep.Flowchartoftheoperationsperjormedinthesingle-sensorobjects.Ifatargetislikelytobeanewmeasurementforanalreadyexistingobject,thetargetislabeled.Insomecasesanobjectcreatesmorethanonemeasurement.Inthiscasemeasurementsareclusteredandlabeledaccordingly.Inthisgatingstepsomeunlikelymeasurementscanbediscarded,e.g.standingobjectsthatmightnotbeofinterestincertainapplications[2],[4].Afterthisstepthedataassociationtakesplace.Thetaskistoassociatethevalidatedtargetstoexistingobjects.Afterthisoperation,allassociatedmeasurementsareusedtoupdatethestateestimationoftheexistingobjects.Ifatargetcouldnotbeassociatedtoanexistingobject,anewobjectiscreated.Ifnomeasurementwasassignedtoanexistingobject,thisobjectiskilled.Objectsmightalsobemergedorsplithere.Outputofthesingle-sensortrackingisalistofobjectsforeachsensor.Attributesofthislistaree.g.distance,lateraldistance,relativelateralvelocityandrelativelongitudinalvelocity.A.Trackingwith77GHzRadarSensorIn[SIand[7]thesingle-sensorradartrackingwaspre-sentedwiththeso-calledinteractingmultiplemodelfilterapproach.TheideaoftheIMMfilterisbrieflyexplainedbyfigure5.TwoKalmanfilterswithdifferentmodelsorparametersetsarerunninginparallelandinteractaccording*'(kIr)Fig.5.Flowchar!oftheIMMfilterprinciples.Fig.6.approach.ResultsoflongitudinalradartrackingusingtheIMMfiltertotheirlikelihood.Ineachtimesteptheactualstatevectorispredictedandinnovatedusingdifferenthypotheses.Theresultingstatevectorisalinearcombinationofthet