Sequentiao Monte Carlo Methods for Multi-Target Filtering with Random Finite Sets

《Sequentiao Monte Carlo Methods for Multi-Target Filtering with Random Finite Sets》由會(huì)員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。

1、I.INTRODUCTIONMulti-targetfilteringisaclassofdynamicSequentialMonteCarlostateestimationproblemsinwhichtheentityofinterestisafinitesetthatisrandominthenumberMethodsforMulti-Targetofelementsaswellasthevaluesofindividualelements[4,5,6].Randomfinitesets(RFSs)areFi

2、lteringwithRandomFinitethereforenaturalrepresentationsofmulti-targetstatesandmulti-targetmeasurements.ThemodellingofSetsmulti-targetdynamicsusingrandomsetsnaturallyleadstoalgorithmswhichincorporatetrackinitiationandtermination,aprocedurethathasmostlybeenperfor

3、medseparatelyintraditionaltrackingBA-NGUVOalgorithms.Moreimportantly,randomsetsprovideUniversityofMelbournearigorousunifiedframeworkfortheseeminglyAustraliaunconnectedsubdisciplinesofdatafusion[15,17,25].SUMEETPALSINGHAlthoughstochasticgeometricalmodels,includ

4、ingCambridgeUniversitydeformabletemplatesandRFSs(orsimplefinitepointU.K.processes)havelongbeenusedbystatisticianstoARNAUDDOUCETdeveloptechniquesforobjectrecognitioninstaticUniversityofBritishColumbiaCanadaimages[2],theirusehasbeenlargelyoverlookedinthedatafusi

5、onandtrackingliteratureuntilrecently[24].TheearliestpublishedworkusingapointRandomfinitesets(RFSs)arenaturalrepresentationsofprocessformalismformulti-targetfilteringappearsmulti-targetstatesandobservationsthatallowmulti-sensortobe[35].Apoint-process-basedfilte

6、rwasalsomulti-targetfilteringtofitintheunifyingrandomsetframeworkproposedin[47]toestimateanunknownbutfixedfordatafusion.Althoughthefoundationhasbeenestablishednumberoftargets.In[32],[33],and[41],ajumpintheformoffinitesetstatistics(FISST),itsrelationshiptoproce

7、sswascombinedwithstochasticdiffusionconventionalprobabilityisnotclear.Furthermore,optimalBayesianmulti-targetfilteringisnotyetpracticalduetoequationsonanon-Euclideanmanifoldtotrackatheinherentcomputationalhurdle.Eventheprobabilitytime-varyingnumberoftargets.Th

8、esameproblemhypothesisdensity(PHD)filter,whichpropagatesonlythefirstwithcontinuousstateevolutionandmarked-pointmoment(orPHD)insteadofthefullmulti-targetposterior,processobservation