外文翻译运动图像和运动矢量检测综述.docx
《外文翻译运动图像和运动矢量检测综述.docx》由会员分享,可在线阅读,更多相关《外文翻译运动图像和运动矢量检测综述.docx(10页珍藏版)》请在冰豆网上搜索。
外文翻译运动图像和运动矢量检测综述
外文文献:
ASURVEYONMOTIONIMAGE
ANDTHESEARCHOFMOTIONVECTOR
Aftermotiondetection,surveillancesystemsgenerallytrackmovingobjectsfromoneframetoanotherinanimagesequence.Thetrackingalgorithmsusuallyhaveconsiderableintersectionwithmotiondetectionduringprocessing.Trackingovertimetypicallyinvolvesmatchingobjectsinconsecutiveframesusingfeaturessuchaspoints,linesorblobs.UsefulmathematicaltoolsfortrackingincludetheKalmanfilter,theCondensationalgorithm,thedynamicBayesiannetwork,thegeodesicmethod,etc.Trackingmethodsaredividedintofourmajorcategories:
region-basedtracking,active-contour-basedtracking,featurebasedtracking,andmodel-basedtracking.Itshouldbepointedoutthatthisclassificationisnotabsoluteinthatalgorithmsfromdifferentcategoriescanbeintegratedtogether.
A.Region-BasedTracking
Region-basedtrackingalgorithmstrackobjectsaccordingtovariationsoftheimageregionscorrespondingtothemovingobjects.Forthesealgorithms,thebackgroundimageismaintaineddynamically,andmotionregionsareusuallydetectedbysubtractingthebackgroundfromthecurrentimage.Wrenetal.exploretheuseofsmallblobfeaturestotrackasinglehumaninanindoorenvironment.Intheirwork,ahumanbodyisconsideredasacombinationofsomeblobsrespectivelyrepresentingvariousbodypartssuchashead,torsoandthefourlimbs.Meanwhile,bothhumanbodyandbackgroundscenearemodeledwithGaussiandistributionsofpixelvalues.Finally,thepixelsbelongingtothehumanbodyareassignedtothedifferentbodypart’sblobsusingthelog-likelihoodmeasure.Therefore,bytrackingeachsmallblob,themovinghumanissuccessfullytracked.Recently,McKennaetal.[11]proposeanadaptivebackgroundsubtractionmethodinwhichcolorandgradientinformationarecombinedtocopewithshadowsandunreliablecolorcuesinmotionsegmentation.Trackingisthenperformedatthreelevelsofabstraction:
regions,people,andgroups.Eachregionhasaboundingboxandregionscanmergeandsplit.Ahumaniscomposedofoneormoreregionsgroupedtogetherundertheconditionofgeometricstructureconstraintsonthehumanbody,andahumangroupconsistsofoneormorepeoplegroupedtogether.Therefore,usingtheregiontrackerandtheindividualcolorappearancemodel,perfecttrackingofmultiplepeopleisachieved,evenduringocclusion.Asfarasregion-basedvehicletrackingisconcerned,therearesometypicalsystemssuchastheCMSmobilizedsystemsupportedbytheFederalHighwayAdministration(FHWA),attheJetPropulsion
Laboratory(JPL),andthePATHsystemdevelopedbytheBerkeleygroup.
Althoughtheyworkwellinscenescontainingonlyafewobjects(suchashighways),region-basedtrackingalgorithmscannotreliablyhandleocclusionbetweenobjects.Furthermore,asthesealgorithmsonlyobtainthetrackingresultsattheregionlevelandareessentiallyproceduresformotiondetection,theoutlineor3-Dposeofobjectscannotbeacquired.(The3-Dposeofanobjectconsistsofthepositionandorientationoftheobject).Accordingly,thesealgorithmscannotsatisfytherequirementforsurveillanceagainstaclutteredbackgroundorwithmultiplemovingobjects.
B.ActiveContour-BasedTracking
Activecontour-basedtrackingalgorithmstrackobjectsbyrepresentingtheiroutlinesasboundingcontoursandupdatingthesecontoursdynamicallyinsuccessiveframes.Thesealgorithmsaimatdirectlyextractingshapesofsubjectsandprovidemoreeffectivedescriptionsofobjectsthanregion-basedalgorithms.Paragiosetal.detectandtrackmultiplemovingobjectsinimagesequencesusingageodesicactivecontourobjectivefunctionandalevelsetformulationscheme.PeterfreundexploresanewactivecontourmodelbasedonaKalmanfilterfortrackingnonrigidmovingtargetssuchaspeopleinspatio-velocityspace.Isardetal.adoptstochasticdifferentialequationstodescribecomplexmotionmodels,andcombinethisapproachwithdeformabletemplatestocopewithpeopletracking.Maliketal.havesuccessfullyappliedactivecontour-basedmethodstovehicletracking.Incontrasttoregion-basedtrackingalgorithms,activecontour-basedalgorithmsdescribeobjectsmoresimplyandmoreeffectivelyandreducecomputationalcomplexity.Evenunderdisturbanceorpartialocclusion,thesealgorithmsmaytrackobjectscontinuously.However,thetrackingprecisionislimitedatthecontourlevel.Therecoveryofthe3-Dposeofanobjectfromitscontourontheimageplaneisademandingproblem.Afurtherdifficultyisthattheactivecontour-basedalgorithmsarehighlysensitivetotheinitializationoftracking,makingitdifficulttostarttrackingautomatically.
C.Feature-BasedTracking
Feature-basedtrackingalgorithmsperformrecognitionandtrackingofobjectsbyextractingelements,clusteringthemintohigherlevelfeaturesandthenmatchingthefeaturesbetweenimages.Feature-basedtrackingalgorithmscanfurtherbeclassifiedintothreesubcategoriesaccordingtothenatureofselectedfeatures:
globalfeature-basedalgorithms,localfeature-basedalgorithms,anddependence-graph-basedalgorithms.
•Thefeaturesusedinglobalfeature-basedalgorithmsincludecentroids,perimeters,areas,someordersofquadraturesandcolors,etc.Polanaetal.provideagoodexampleofglobalfeature-basedtracking.Apersonisboundedwitharectangularboxwhosecentroidisselectedasthefeaturefortracking.Evenwhenocclusionhappensbetweentwopersonsduringtracking,aslongasthevelocityofthecentroidscanbedistinguishedeffectively,trackingisstillsuccessful.
•Thefeaturesusedinlocalfeature-basedalgorithmsincludelinesegments,curvesegments,andcornervertices,etc.
•Thefeaturesusedindependence-graph-basedalgorithmsincludeavarietyofdistancesandgeometricrelationsbetweenfeatures.
Theabovethreemethodscanbecombined.IntherecentworkofJangetal.[34],anactivetemplatethatcharacterizesregionalandstructuralfeaturesofanobjectisbuiltdynamicallybasedontheinformationofshape,texture,color,andedgefeaturesoftheregion.UsingmotionestimationbasedonaKalmanfilter,thetrackingofanonrigidmovingobjectissuccessfullyperformedbyminimizingafeatureenergyfunctionduringthematchingprocess.
Ingeneral,astheyoperateon2-Dimageplanes,feature-basedtrackingalgorithmscanadaptsuccessfullyandrapidlytoallowreal-timeprocessingandtrackingofmultipleobjectswhicharerequiredinheavythruwayscenes,etc.However,dependence-graph-basedalgorithmscannotbeusedinreal-timetrackingbecausetheyneedtime-consumingsearchingandmatchingofgraphs.Feature-basedtrackingalgorithmscanhandlepartialocclusionbyusinginformationonobjectmotion,localfeaturesanddependencegraphs.However,thereareseveralseriousdeficienciesinfeature-basedtrackingalgorithms.
•Therecognitionrateofobjectsbasedon2-Dimagefeaturesislow,becauseofthenonlineardistortionduringperspectiveprojectionandtheimagevariationswiththeviewpoint’smovement.
•Thesealgorithmsaregenerallyunabletorecover3-Dposeofobjects.
•Thestabilityofdealingeffectivelywithocclusion,overlappingandinterferenceofunrelatedstructuresisgenerallypoor.
D.Model-BasedTracking
Model-basedtrackingalgorithmstrackobjectsbymatchingprojectedobjectmodels,producedwithpriorknowledge,toimagedata.Themodelsareusuallyconstructedoff-linewithmanualmeasurement,CADtoolsorcomputervisiontechniques.Asmodel-basedrigidobjecttrackingandmodel-basednorigidobjecttrackingarequitedifferent,wereviewseparatelymodel-basedhumanbodytracking(norigidobjecttracking)andmodel-basedvehicletracking(rigidobjecttracking).
1.Model-BasedHumanBodyTracking:
Thegeneralapproachformodel-basedhumanbodytracingisknownasanalysis-by-synthesis,anditisusedinapredict-match-updatestyle.Firstly,theposeofthemodelforthenextframeispredictedaccordingtopriorknowledgeandtrackinghistory.Then,thepredictedmodelissynthesizedandprojectedintotheimageplaneforcomparisonwiththeimagedata.Aspecificposeevaluationfunctionisneededtomeasurethesimilaritybetweentheprojectedmodelandtheimagedata.Accordingtodifferentsearchstrategies,thisisdoneeitherrecursivelyorusingsamplingtechniquesuntilthecorrectposeisfinallyfoundandisusedtoupdatethemodel.Poseestimationinthefirstframeneedstobehandledspecially.Generally,model-basedhumanbodytrackinginvolvesthreemainissues:
•Constructionofhumanbodymodels;
•Representationofpriorknowledgeofmotionmodelsandmotionconstraints;
•Predictionandsearchstrategies.Previousworkonthesethreeissuesisbrieflyandrespectivelyreviewedasfollows.
A.Humanbodymodels:
Constructionofhumanbodymodelsisthebaseofmodel-basedhumanbodytracking.Generally,themorecomplexahumanbodymodel,themoreaccuratethetrackingresults,butthemoreexpensivethecomputation.Traditionally,thegeometricstructureofhumanbodycanberepresentedinthefollowingfourstyles.
•Stickfigure.Theessenceofhumanmotionistypicallycontainedinthemovementsofthetorso,theheadandthefourlimbs,sothestick-figuremethodistorepresentthepartsofahumanbodyassticksandlinkthestickswithjoints.Karaulovaetal.useastickfigurerepresentationtobuildanovelhierarchicalmodelofhumandynamicsencodedusinghiddenMarkovmodels(HMMs),andrealizeview-independenttrackingofahumanbodyinmonocularimagesequences.
•2-Dcontour.Thiskindofhumanbodymodelisdirectlyrelevanttohumanbodyprojectionsinanimageplane