recognisingyahoo.cn

ANEWFRAMEWORKOFMOVINGTARGETDETECTIONANDTRACKINGFORUAVVIDEOAPPLICATIONWenshuaiYua,*,XuchuYub,PengqiangZhang,JunZhouaInstituteofSurveyingandMapping,450052,Zhengzhou,Henan,China-ws_yu@yahoo.
cnbInstituteofSurveyingandMapping,450052,Zhengzhou,Henan,China-xc_yu@yahoo.
comWGS,WGIII/5KEYWORDS:ImageProcessing,ComputerVision,MotionCompensation,MotionDetection,ObjectTracking,ProcessModeling,UAVVideoABSTRACT:Unmannedaerialvehicleisanewplatformforremotesensing,andtheprimarysensorofitisvideocamera.
Video,alsocouldbecalleddynamicimageisthemostimportantdataformatwhichobtainedbyunmannedaerialvehicle.
ThecombinationofvideodataandUAVprovidesanovelremotesensingpattern.
Movingtargetdetectionandtrackingisanimportanttechniqueofvideoprocessingforitshugepotentialinmilitaryandotherapplications.
Thetechniquealwayscontainsthreebasicparts:motioncompensation,motiondetectionandobjecttracking.
Eachpartadoptskindsoftechnicalmethodstosolvetheproblemsinrespectivefields.
Thepaper,basedontheanalysisofthealgorithmsrelatedtothetechnology,presentsanewframeworkofit.
Differentfromothermovingtargetdetectionandtrackingframeworks,theframeworkperformsaparallelprocessingamongthethreesectionsbyincludingcollaborationcontrolanddatacapturemodules.
Comparingwithotherframeworks,itismoresuitabletotheUAVapplications,becauseofitsadvantagessuchastransferringparametersinsteadofrealdataandofferinginterfacetouserorexteriorsystem.
*Correspondingauthor.
Tel.
:+86-13526657654;E-mailaddress:.
ws_yu@yahoo.
cn.
1.
INTRODUCTIONUnmannedAerialVehicle(UAV)isanewdevelopingremotesensingplatform,anddifferentfromotherplatforms,forexamplesatelliteorairplane,itcarriesvideosensors.
SovideodataisthemaininformationgotbyUAV.
Videocouldbeinterpretedasdynamicimage,anddissimilartostaticimage,itcanreflectmotioninformationthroughthechangingofgray-level.
AnimportantresearchfieldofvideoprocessingforUAVapplicationismovingtargetdetectionandtracking.
Inactualenvironment,themovingtargetscouldbevehicles,peopleoraircrafts,andinsomespecialconditions,thesetargetsmightbeinterestingandvaluable.
Buttheproblemthatdetectingthetargetsfromthecomplicatedbackgroundandtrackingthemsuccessivelyisatoughwork.
Theremanytechniquemethodsonmovingtargetdetectionandtracking.
Mostofthemanalysedtheproblemundertheconditionofstaticbackground,forthestillnessofbackgroundmakesthedetectionandtrackingcomparativelyeasier,andthesekindsofmethodcanbeusedinsomeapplicationssuchassafetymonitoring.
Contrastingtothem,itismuchmoredifficultfortargetdetectionandtrackingwithmovingbackground.
EspeciallyforUAVvideodatawhosebackgroundchangingrapidlyandalwayshascomplextexturecharacteristic,itisreallyachallengingtasktosolvethetechnicalproblem.
FormovingtargetdetectionandtrackingusingUAVvideo,aratherreasonabletechnicalapproachisadoptedwidely.
Firstly,inordertocompensatethebackgroundmotioncausedbymovementofcamera,stabilizingthebackgroundthroughtheframe-to-frameregistrationofvideoimagesequencewouldbetakenasapreconditionofdetectionandtracking.
Asignificantproductthepanoramicimageisbuiltinthesameprocess.
Secondly,basingonthestabilizationofbackgroundandemployingpropermethods,thenextoperationisseparatingthetargetimagefromthebackgroundtorealizedetectionofmovingtarget.
Finally,movingtargettrackingislocatingtheobjectinimagebymeansofmodelingthetargetaccordingtotarget'sfeaturepropertyandchoosingappropriatetrackingmethod.
Accordingtothetechnicalapproachmentionedabove,thetechniquecanbedividedintothreesections:motioncompensation,motiondetectionandobjecttracking.
Italwaystakesthethreepartsasaserialcourseandimplementsthemoneafteranotherinaprocessing.
Actually,fortherearemutualactivitiesbetweendifferentsectionsofthetechnique,itisnotnecessarytoprocessthetechnologyorderly,whichmeansexecutingitstepbystep.
Soitnotonlyneedsaframeworktointegratealltheseparts,butalsorequirestheframeworkmoreeffectiveandpractical.
2.
MOTIONCOMPENSATIONMotioncompensationisthebasicpartofthetechnique,especiallyformovingbackgroundvideo.
Itestimatestheego-motionofcameraandcompensatesthebackgroundmotionofimage,andthroughthisway,itmakesthemovingobjectsmoreobviousandthedetectionoftargeteasier.
Therearetwokindsofapproachesadopted,oneisfeature-basedmethods,andtheotherisflow-basedmethods.
Thoughthelatteronehasrigoroustheoryfoundation,theformeroneismorepopular.
Feature-basedmethodsextractfeaturesandmatchthembetweenimageframestofittheglobalmotionmodelofvideoimagesequence.
Featureextractionandmatchingarepreparedforimageregistration.
Theimageregistrationthatimplementsframe-to-frameregistrationofthevideoimagesequenceisthekeypointofmotioncompensation.
Theresultofimageregistrationcouldbeusedintwodirections,imagestabilizationandimagemosaicking.
Formercanrestrainthemovingbackgroundandfacilitatethedetectingofmovingtarget,andlattercanupdatethelocalimage(alwaysexpresswiththeortho-image)andhelptoformthetrajectoryoftrackedobject.
2.
1FeatureExtractionandMatchingInfeatureextraction,choosingarightkindoffeatureshouldbeconsideredforonething.
Thefeaturecouldbepoint,lineorsurface.
Ithasbeenproventhatcornerfeatureisrobustandeasytooperate.
Harrisoperator(Harrisetal.
,1988)isatypicalcornerdetector,anditsprincipleisthatrecognisingthefeaturesbyjudgingthedifferenceofgray-level'schangewhilemovingthesearchwindow.
Detectingresultsoftwoseriesframesshowninfigure1,andthereisgoodcoherencebetweenthetwo,soitshouldbethoughtthattheoperatorhasastableperformanceandtheresultscouldbetakenastheinputofmatching.
Afterextractingthefeatures,acoarsematchingwouldbemadetogetapproximatematchingresults,andthiscourseisrealizedbymeasuringthesimilarityofcorrespondingfeatures.
Becausetherearemanymismatchesintheapproximateresultsandtheycannotmeettherequirementsofregistration,soithastoimplementafinematchingtoremovethemismatches.
AsuitablewaytokeepinliersiscombiningofepipolargeometryandRANSACalgorithm.
Epipolargeometryoffersamodel—fundamentalmatrixtothematching,causethetwoviewsshouldsatisfytheepipolarrestrictioninstereovision.
RANSAC—randomsampleconsensusalgorithm(Fischleretal.
,1981)isanonlinearalgorithm.
FittingdatamodelwithRANSACmaximallyrestrainstheimpactofoutliers,andreducesthecomputationtoacertainextent.
Thefinematchingisfittingthefundamentalmatrixthroughiterationcomputingandidentifyingmostoftheoutliers.
Figure2presentstheresultsofmatchingaftereliminatingwrongcorrespondencesfromthecandidatematcheswhichgotfromthecoarsematching.
Itcanbeseenthatthoughbulkofmismatcheshavebeenremoved,therestillafewincorrectcorrespondencesremain.
2.
2ImageStabilizationImagestabilizationiscompensationofunwantedmotioninimagesequences.
Thematterofimagestabilizationisimageregistration.
Thetransformationmodelofimageregistrationisnotcomplicate.
Ausualchoiceisaffinetransformationorprojectivetransformation.
Figure3.
ThecomparationofdifferenceresultsbeforeandafterimageregistrationThenormalmodeforregistrationiscalculatingtheparametersofthemodelusingcorrespondingpoints.
Whethertheprecisionofimageregistrationisgoodornotdependsontheresultsofmatching.
Soimagestabilizationcouldbedonebycomputingtheregistrationparameterswiththeoutputsoffinematchingandrectifyingthepreparedframetoreferenceframe.
Inordertooptimizetheresultofregistration,repeatingthecourseuntiltheaccuracyofregistrationgoodenough.
Figure3showsthecomparationofdifferenceresultsbeforeandafterimageregistration.
Theleftoneisthedifferenceresultpreviousregistration.
Exceptsomeregionswithsametextures,mostofthebackgroundimagecannotbesubtracted,especiallysomeobviousobjectsandlinearfeatures.
Therightoneisthedifferenceresultafterimageregistration.
Thoughthereareobjectsedgesstilldistinct,majorityofbackgroundimagegotbetterelimination.
Figure1.
DetectingresultsusingHarriscorneroperator2.
3ImageMosaickingMosaickingofvideoimagesequenceisrectifyingallframestothereferenceframeandpiecingthemtogetherasapanoramicimage.
Thereferenceframemaybethefirstframeorachosenone.
Akeystepforthegenerationofpanoramaisimageregistration.
Itisunavoidableaccumulateregistrationerrorsduringaligningtheimagesequences.
Theaccumulationoferrorscouldinducemisalignmentofadjoiningframes.
Toresolvetheproblem,therearemanymethodshavebeentried,suchasrefiningregistrationandintroducingreferencedata.
AnUAVvideoimagemosaickingisillustratedinfigure4,andtherearesomepiecingseamsforregistrationerrors.
Figure2.
OverlayoftwosuccessiveframesaftereliminatingwrongcorrespondenceswithRANSAC3.
MOTIONDETECTIONThecompensationhasreducedtheimpactofbackgroundmotion,buttherearestillsomeinfluencesofitremaininthestabilizedimage.
Motiondetectiondividesthevideoimageintotargetandbackgroundwhetheritismovingornot.
Therearemanyprocessingmethodsintroducedintomotiondetection,andthecommonpointofthemistheusingofmotioninformation.
Forstaticbackground,itusuallyprocessesonthebackground,suchasbackgroundmodelingmethod.
Formovingbackground,itassumesthedynamicimagejusthastargetandbackgroundtwopartitions,andiftherearemorethanonetargetinthevideo,itwillsegmenttheimageintonumbersofpartitionscorrespondingtothetargets,andinsomemethodsitsetsthetargetsondifferentlayersinordertomaketheprocessmuchfaster.
Theprimaryinformationfordetectingismotioninformation,ortheintensitychangesbetweenadjacentvideoimageframes.
3.
1MotionDetectionForvideoimagecapturedbymovingcamera,thebackgroundmotioncan'tbecounteractedabsolutelythroughimagestabilization.
Itmaynoteffectiveenoughtodetectthemovingtargetbyrestrainingthemovementofbackground.
Alltheimageinformationcouldbeclassifiedintothreekinds:target,backgroundandnoise.
Differentclassescorrespondtodifferentmotionfieldsindynamicimage.
Ifweknowtheclasscharacteristicsofpoints,wecanusethemtofittheparametricsetsofdifferentmotionregions.
Contrarily,ifweknowtheparametersofmotionvectors,wecoulddividethepixelsintodifferentfieldsaccordingmotioninformation.
Inmostofcases,bothofthecharacteristicsandparametersareunknown.
Theclusteringofimagepixelsisaprobabilityquestion.
AtypicalsolutionformotionclassificationisunitingthemixtureprobabilitymodelandEM—ExpectationMaximumalgorithm(Weissetal.
,1996).
Inpractice,itcanmakeahypothesisthattherearetwolayersinthedynamicimage,backgroundlayerandtargetlayer.
Afterimagestabilization,calculatingthemotionvectorsofallpixelsandassumingthattheflowvectorsoftargetlayerislargerthantheonesofbackgroundlayertoestimatetheweightsofmixturemodelwithiteratedcomputation.
Itwillhavethetargetdetecteduntiltheiterationconvergence.
Theparametersofimageregistrationcouldbetheinitialvaluesofiteration.
Figure5presentsadetectionresultforonevehicletargetinthreeframes.
3.
2MotionSegmentationmthesegmentationwiththeopticalflowformationonly.
enodesinthiswindowwhenconstructingtheweightedgraph.
4.
OBJECTTRACKINGFigure5.
AmotiondetectionresultwithmixturemodelndEMMotionsegmentationisakindofvideosegmentation,becauseitpartitionsvideoorimagesequenceintospatio-temporalregionsbasingonmotioninformation.
Therefore,itisessentiallysameasthemotiondetection.
Generally,motionsegmentationhastwobasicclassesthatopticalflowsegmentationmethodsanddirectmethods(Boviketal.
,2005).
Inperfectcases,therearejusttwokindsofopticalflowassociatedwiththemovementsofbackgroundandtarget.
However,opticalflowisnotanexactreflectionofmotionfieldbutanexplanationofilluminationchange.
Therefore,itisnotrigoroustoperforinAusuallyadoptionisgroupinginmotionfeaturespacetorealizethesegmentation.
Howtosettherelationbetweenclusteringanddynamicimageisanotherquestion.
Themethodofgraphtheoryisanaturalsolutionformotionsegmentation.
Pixelsinimagesequencecouldbetakenasthenodesofgraph,andifwepartitionthegraph,accordingmotionfeatures,maysegmenttheimageatthesametime.
Edgetheweightmeansthesimilarityoffeaturesbetweenthetwonodeswhichconnectedbyit.
Inmotionsegmentation,thissimilaritymeasurementisthemotionfeaturevectorofeachpixel.
Thegraphisnotconstructedinoneimageframe.
Itshouldconnectallthenodesinaspatiotemporalregion,andtheregionmayacrossseveralframes.
Aftertheconstructionoftheweightedgraph,itcouldsegmentthevideoimagesequenceusingbynormalizedcutmethod(Shietal.
,1998).
Inordertoreducethecomplicationofcomputing,aneffectivesolutionissubsamplingtheimagesequencebysettingspatiotemporalwindowthatjustconnectthAfterdetectingthelocationoftargetinimage,objecttrackingwillpersistentlylockthepositionoftargetduringaperiod.
Thebasicideaofobjecttrackingismodelingtheobjectaccordingtoobject'sfeaturecharacteristicpropertyandchoosingappropriatetrackingmethod.
Differentformmotiondetectionemphasizingonaccuracy,objecttrackingcouldn'tabidetakingtoomuchtimeoncomputingandneedsgivingattentiononbothprocessingspeedandprecision,soithastoabstractthetargetthroughfeatureextractionandobjectmodeling.
Simplythefeaturesusedcouldbeshape,size,directionandvelocityofthemovingobject,andcomplicatedlyitcouldbefeaturepointsset,colorspaceandsoon.
Combiningwithrespectivetechnicalapproach,itwillrealizethetargettracking.
Theessenceofobjectmodelingistryingtodefinethetargetuniquely,andinaFigure4.
ApanoramicimagemosaicedbyUAVvideoimagesequencesingletargettrackingitonlyneedtodependononefeatureproperty,butinmulti-targettrackingitmayneedaintegrationofdifferentkindsoffeaturesfordirectingatpropertarget,anditalsocouldusingsomesuitableways,suchasfiltermethodsrmulti-target.
4.
1ObjectModelingirectly,ortransformthemintootherrmssuchastemplates.
singmulti-featuresmodelandupdatingthemodel4.
2ObjectTrackingingintothematchingofpointsets(Huttenlocheretl.
,1993).
epeatseprocessuntilthefilterisstable(Forsythetal.
,2003).
irbornevideousingMean-shiftmethodisowninFigure6.
5.
SYSTEMFRAMEWORKandetrackingresultcanacceleratethedetectionprocessing.
them,anditprovidesinterfacetouserandexteriorstem.
foObjectmodellingisarepresentationofobject,inotherwordsitutilizesonefeaturecharacteristicorthecombinationoffeaturestoexpresstheobject.
Theobject'sfeaturecouldbecontour,shape,color,position,texture,velocityandsoforth.
Themorefeaturesincluded,theeasiertoidentifytheobject.
Butthecombiningfeatureswillincreaseburdenofprocessinganddemandcompositemethods.
Toconstructthemodelofobject,wecanusethefeaturesdfoFeaturesoftheobjectmaychangeduringthecourseoftracking,soitrequiresthatthemodelshouldbeadaptivetothechangingorotherinfluences,forexampleocclusionandunexpectedmovement.
Thisisconsideredastherobustnessofmodel.
Therearemanywaystomakethemodelmorestable,includinguovertime.
Usingpriorinformationthatformsthemodelofobject,trackerpredictstheobject'spositioninsuccedentframes.
Correspondingtodifferentmodels,objecttrackinghasdifferentmethods.
Objecttrackingmethodsattempttoascertainthecoherentrelationsoffeatureinformationbetweenframes,andthestrategyofitisnomorethansearchingandmatching.
Hausdorffdistanceisavalidmeasurementforshapeandtexturefeaturesoftheobject.
Itcancreatesparsepointsetswithfeaturedetectorsinimages,andthepointsetofimageregionlabelledastheobjectistheobject'smodelforHausdorffmeasurement.
Itisabletotacklethedeformationofobject,becauseitdescribesthecontourandtextureoftheobjectwithbulkofpoints.
Takingthemeasurementandthemodel,ittranslatesobjectlocataMotionisakindofstate.
Atypicalmotionstatevectoriscomposedoftheobject'sposition,velocityandaccelerationalongeachdirection.
Ifthepriorandcurrentstatesareknown,theposteriorstatewillbepredicted.
Itisfeasibletoresolvetheproblemofobjecttrackingbystateestimationmeans.
Kalmanfilterisoneofthestatespacemethods.
Todefineit,theKalmanfilterisabatchofmathematicequationsthatsolvestheleast-squaresquestionrecursively.
Itpredictsthevaluesofcurrentstateutilizingtheestimationvaluesofformerstateandtheobservationvaluesofcurrentstate,executingtheprocedurerecurrentlyuntilthevaluesofeverystateestimated.
Togettheestimationvaluesofeachstate,allthepreviousobservationvalueshavebeeninvolved.
Forobjecttracking,thestateequationisthemodelofobjectinKalmanfilter,anditdescribesthetransferofstates.
Theobservationisthepositionofobject,andthestatevectorlikementionedabovecontainsposition,velocityandacceleration.
PuttingthepositionsofobjectdetectedininitialframesintotheobservationequationofKalmanfilterandtakingtheaccuratepositionsastheinitialvalueofstatevariant,itcomparestheoutputoffilteringwithpreciseresulttotestifythecorrectnessofinitialinput.
ItrthMean-shiftalgorithmisanapproachthatsearchesthemaximumofprobabilitydensityalongitsgradientdirection,aswellasaneffectivemethodofstatisticaliteration.
ObjecttrackingwithMean-shiftalgorithmisanotherclassoftechniquethatlocatesthetargetbymodelingandmatchingit.
Boththemodelingandmatchingareperformedinafeaturespacesuchascolorspaceandscalespace.
Themodeofitisusingtherelevantsimilaritymeasurementtosearchthebestmatch.
TheobjecttrackingbasingonMean-shiftalgorithmmainlyprocessesonthecolorfeature.
Choosinganimageregionasthereferenceobjectmodel,itwillquantizethecolorfeaturespace,andthebinsofthequantizedspacerepresenttheclassesofcolorfeature.
Eachpixelofthemodelcancorrespondstoaclassandabininthespace,andthemodelcanbedescribedbyitsprobabilitydensityfunctioninthefeaturespace.
InsteadofPDF(probabilitydensityfunction),ittakesthekernelfunctionasthesimilarityfunctiontoconquerthelostofspatialinformation.
Anotherreasonforusingkernelfunctionissmoothingthesimilaritymeasurementtoensuretheiterationconvergetotheoptimizedsolutionduringsearch(Comaniciuetal.
,2003).
AnobjecttrackingresultofashFigure6.
AnobjecttrackingresultofairbornevideousingMean-shiftmethodTothetechnicalapproachesanalysedabove,itneedsaframeworktointegrateallthesemethods.
Forthetechniqueofmovingtargetdetectionandtrackingdividedintothreeparts,eachpartwouldbeanisolatedmoduleforitsindependentfunctioninapplicablesystem.
Therefore,theprocessingisinandbetweendifferentmodules.
Therearemanysystemsemployaseriesprocedure.
Compensationcomesfirst,thenextisdetection,andtrackingputonthelast.
Thereasonofthatisanteriormodulealwaysbetakenasthepreconditionofposteriormodule,andresultsofeachonecouldbeinputsofthenextone.
However,thiskindofsystemisnotconsideringtheinteractionsbetweendifferentmodules.
Forexample,theresultofsegmentationcanbetheinitialvalueofcompensation,thAsshowninthefigure7,distinguishingfromtraditionaltechniqueframework,thepresentedsystemframeworkintroducestwomoremodules,whicharedatacaptureandcollaborationcontrol.
Datacapturemodulegetsthevideoimagedataandsamplesitintoimagesequence,andthenitwilldistributethemtoanotherthreemodulesthatarethecentralpartsofthesystem.
Thethreemodulesimplementaparallelprocessing,andthiswilllowerthecostoftime.
Aftertheinteriorcomputing,theytransfertheoutputsthatalwaysinthemannerofparameterstocollaborationcontrolmodule.
Thecontrolmodulemanagesalltheothermodulesbysendingorderstosyeobjectbyutilizingmethodscorrespondingtothemodelofit.
computationtomeettherequirementofreal-timeapplication.
insteadofrealdatatominimizesthetransmissionbandwidth.
ontrolmoduletovaluatethemethodsormakeimprovement.
ethods,anotheruseofthisframeworkistestingthenewbornethods.
alUAVstemcomposesofaircraftandgroundcontrolstation,andthewirelesscommunication.
Oamework,constructthetestbedsystemtotesttheperformanceoftechnicalmethodsandsetthes(2)EmbeddingthefunctionalmodulesintotheUAVsystemndimprovingthemtomeetthepracticalrequirements.
ephens,M.
,1988.
ACombinedCornerandEdgeetector.
FourthAlveyVisionConference,ManchesterUK,FittingwithApplicationstoImagenalysisandAutomatedCartography.
Communicationsofthegmentation:incorporatingspatialcoherenceandtimatingthenumberofmodel.
InProc.
IEEEConf.
onCVPR,lik,J.
,1998.
MotionSegmentationandTrackingsingNormalizedCuts.
Proc.
Int'lConf.
ComputerVision,pp.
UsingtheHausdorffDistance.
IEEEansactionsonPatternAnalysisandMachineIntelligence,orsyth,D.
A.
,Ponce,J.
,2003.
ComputerVision:AModern2003.
Kernel-BasedObjectTracking.
IEEEtransactionsonPatternAnalysisandMachineIntelligence,25(5):564-577mm6.
CONCLUSIONOnthebasisofanalyzingthefunctionalpartsthatmotioncompensation,motiondetectionandobjecttrackingandthecorrespondingtechnicalmethodsofmovingtargetdetectionandtracking,wepresentedanewframeworkforthetechnique.
Werecognizethatalthoughthereareconnectionsbetweendifferentsectionsofthetechnology,aserialprocessingofthemisdispensable.
Werealizedaparallelcomputationofthethreepartsbyaddingcontrolandcapturemodules.
Thedesignoftheframeworkfacilitatesthespatialseparationofsystemandreducesthedatastreamtransferredbetweendifferentmodules.
ThisismeaningfultoUAVapplication.
BecauseatypicsydatatransferringdependsonFigure7.
Movingtargetdetectionandtrackingframeworkurfurtherworkincludes:(1)AccordingtothefrFigure8illustratesthemainfunctionalmodulesofthesystem.
Motioncompensationhasimagemosaickingandimageregistrationtwoparallelsub-modules.
Imagemosaickingthatcouldcombinewithotherdatamosaicstheimagesequence,andimageregistrationcalculatesregistrationparametersoropticalflowvectors.
Motiondetectionincludesbackgroundsubtractionandtargetdetectiontwoserialsub-modules.
Backgroundsubtractionrestrainsthemovementofbackgroundusingtheparametersorthevectors,andtargetdetectionextractstargetfromthecompensatedbackground.
Objecttrackingcontainstwoserialsub-modulesthatareobjectmodelingandobjecttracking.
Objectmodellingconstructsthemodelofobjectwithitsfeatures.
Objecttrackingrealizesthesuccessivelocatingoftandardforevaluation.
aREFERENCESHarris,C.
,StDpp.
147-151Fischler,M.
A.
,Bolles,R.
C.
,1981.
RandomSampleConsensus:AParadigmforModelthTheadvantagesofthisframeworklistedasbelow:sAACM,24(6),pp.
381-395Weiss,Y.
,Adelson,E.
H.
,1996.
Aunifiedmixtureframeworkformotionseepp.
321-326Alan,B.
,2005.
HandbookofImageandVideoProcessing.
Elsevier,pp.
474-485Shi,J.
,MaFigure8.
MainfunctionalmodulesofthesystemU(1)Parallelprocessingreducesthe1154-1160Huttenlocher,D.
P.
,Noh,J.
J.
,Rucklidge,W.
J.
,1993.
ComparingImages(2)Transferringkindsofparameterstr(3)Usersandexteriorsystemscanconductandmonitorthemodulesthroughtheinterfacesofferedbyc15(9),pp.
850-863FeApproach.
PrenticeHall,pp.
380-396Comaniciu,D.
,Ramesh,V.
,Meer,P.
,Movingtargetdetectionandtrackingisadevelopingtechnique,andmanytechnicalmethodswillbeinventedandintroducedforitinfuture.
Thoughthemethodsmaybediverseinformsandbasedtheories,theyhaveanidenticalpurposeandconformtoaregularsystemframework.
Besidesintegratingtheexisting