基于改进粒子滤波器目标跟踪算法的研究

安徽大学硕士学位论文基于改进粒子滤波器目标跟踪算法的研究姓名：钱翔申请学位级别：硕士专业：信号与信息处理指导教师：李新华2010-05I――Kalmanfilter,KFExtendedKalmanfilter,EKFUnscentedKalmanfilter,UKFParticlefilter,PFKFEKFUKFEKFPFPFUKFEKFEKFKFRGBCMYKHSVHSVIIHSVAbstractIIIAbstractTargettrackingisthecorecomponentsofintelligentsystemstodeterminelocation,movementandidentitytargets,whichiswidelyusedinthefieldofvideosurveillance,securitysystemsandintelligenttransportationsystem.Itisreallyatoughjobtorealizethissystemandfindawidelyusedandhighrobusttrackingalgorithmduetocameramovement,targetinstability,complexityofbackgroundandmovingsimilarity.Itishardtofind.Kalmanfilteralgorithm,proposedbyMr.Kalmanmorethan40yearsago,isthebestwaytosolvetheprobleminthelinearGaussianenvironment.However,inordertomeettechnologyandapplicationneeds,therehasbeenanemergenceofstudyingthenonlinearnon-Gaussianfilteringalgorithmrecentyears.First,thisthesisintroducedthecommontargettrackingfilteralgorithmproposedrecently,suchasKalmanFilter(KF),extendedKalmanfilter(EKF),UnscentedFilter(UKF)andparticlefilter(PF)algorithm.Simpleandelegant,KFisthebestrecursiveBayesianestimatorinlinearGaussianenvironment.ByusingTaylorseries,EKFtransformsnonlinearproblemintolinearspace,thenusingKalmanfiltertoestimatetheresultstoachievethefirstorderaccuracy.Throughthefixedsamplesettoapproximatetheprobabilitydistributionofthestate,UKFisbetterthantheEKFonprecisionandquantity.Nonetheless,asusingGaussianposteriorprobabilitydensitytoapproximatethesystemstate,itispoortoperformincomplexenvironment.Particlefilter(PF)isaBayesianfilteringadoptedbyMonteCarlosamplingmethod.Thecomplextargetstatedistributionisexpressedasasetofweights(calledparticle)inthisfilter.Byfindingthelargestweightparticlesintheparticlefiltertodeterminethemostlikelytargethasbeenprovedasthebestwaytotracktargetinacomplexenvironment.Bymeasuringthenonlinearmodel(tangent),thisthesisdemonstratedthattheparticlefilterhasthemostoutstandingperformancedealingwiththenonlinearsituations,andUKFhassuperiorperformancethanEKF,EKFisbetterthanKF,whichisidenticaltothetheoreticalanalysis.Second,itisatoughjobtoselectthecharacteristicsoftargetsintargettrackingIVsystem.Iftargetshavemorefeatures,trackingaccuracycouldbeeffectivelyimproved,however,computingquantityandcalculationtimewouldalsoincreasing.Itisimperativeforustotakecompromiseofreal-timeandaccuracy.Ashighstabilityandlowcomputationalcharacteristics,colorhistogramfeaturearebecomingamainfeaturetodescribetargets.ThisthesisintroducedRGBspace,CMYKspace,HSVspace.Allthesespaces,HSVspaceismoresuitableforhumantoperceivecolor.Also,thismodelhasgoodlinearscalabilityadvantages.However,singlecolorhistogramissensitivetothebackgroundilluminationandwhat’smore,trackingaccuracycouldreducedsignificantlywhenthissystemisinterferedwithsimilarcolorobjects.Bycontract,asobjectsmomentsfeaturehasstructuralcharacteristics,whichhavethepropertiesoftranslation,rotationandscaleinvariant,etc.Itiswidelyusedinimagematchingandgesturerecognition.Finally,bycombiningbothpropertiesoftargetcolorandmomentinvariantthisthesispresentedanimprovedparticlefilterbasedonmethod,thecharacteristicofthecolorhistogramiscarriedoutinHSVspace.Theweightsoftheparticlearedeterminedbyapplicationenvironment.Further,bydeterminedEuclideandistancepropertiesintheprocessofreplacement,poorqualityparticleswerewashedout,reliabilityofparticleswereincreasedandtheimpactofnoiseswerereduced.Theexperimentalresultshavebeendemonstratedthatthismethodamelioratedtheinterferenceimmunityofthesinglecolorpropertyfortrackingtarget.Inaddition,thismethodalsoimprovedthetrackingaccuracyandrobustnesswhilenotaffectingthereal-timecharacteristics.Keywords:TargettrackingParticlefilteringColorhistogramInvariantmomentCombiningfeatures11.180%[1][29]Wax1955[2]Sittler1964701975Bar-ShalomKalman1-11-1Figure1-1Structureofatypicalmonitoringsystem:2[3][4][6]1.2[44][45]132[7][31]3[36]4[10][15][16][43][17][18]4MPEG4H.264MPEG[39]1.3KalmanKalmanfilter,KF[8][9]ExtendedKalmanfilter,EKF[26][27]UnscentedKalmanfilter,UKF[23][24]Particlefilter,PF[37][38][42][30]51-1I(t)=g(Y(t),K(t),t)1-1Y(t)nmI(t)gmK(t)[t0,t]Y(t)I=I(β)t0tyΛ(I)Y(t)yΛ(I)Y(t)[5][12]1960ANewApproachtoLinearFilteringandPredictionProblems[25]6EKFUKF1.4[17][18]MicrosoftVisualStudio6.0IntelOpenCVGSL72.1KF2.1.1KFxnR∈111kkkkXAxBuw−−−=++2-1mzR∈kKKzHxv=+2-2Kvkw:()pw:()0,NQ()()0,PvNR:QR.1ku−1kw−nn×AK-1KA2mn×HKxkz2.1.2nkxR−∧∈-∧KKnkxR∧∈kzKkkkexx−∧−−=−2-3kkkexx∧−=−2-4TkkkpEee−−−=2-58Tkkkpee=2-62-7kx−∧kzkHx−∧kx∧kkkkxxKzHx−−∧∧∧=+−2-72-7kkzHx−∧−2-7mn×K2-6K2-7keke2-62-6kpKKK()1TTTkkkkTkpHKpHHpHRHpHR−−−−−=+=+2-82-8RKR1limKKRKH−→∞=2-9kp−Kkp−lim0kKpK−→∞=2-10KRkzkzkHx−∧kp−kzkzkHx−∧2.1.39111kkkTkkxAxBupApAQ−ΛΛ−−−−=+=+2-11kkkKkxxKzHx−−∧∧∧=+−2-121()TTTkkkkTkPHKPHHPHRHPHR−−−−−=+=+2-13(1)kkkPKHP−=−2-14kKkz2-12-1Figure2-1Kalmanfiltermodel10EKFUKF2.2EKF2.2.1EKFEKFEKFEKF2.2.2EKF2-152-16xn+1=1(,()())fnxngn+2-15yn+1=2(,()())hnxngn+2-162-172-18fn+1/n=(/)(,)xxnnfnxx=∂∂2-1711hn=(/1)(,)xxnnhnxx=−∂∂2-182-172-18f()h()2-17f(n+1/n)ijf(n,x)ixjh(n)iJh(n,x)ixj2-17x(n|n)2-18x(n|n-1)x(n|n)x(n|n-1)f(n+1/n)h(