2003_Detecting Shadows_Algorithms and Evaluation

阴影检测经典论文

918IEEETRANSACTIONSONPATTERNANALYSISANDMACHINEINTELLIGENCE,VOL.25,NO.7,JULY2003

DetectingMovingShadows:AlgorithmsandEvaluation

AndreaPrati,Member,IEEE,IvanaMikic,Member,IEEE,

MohanM.Trivedi,Member,IEEE,and

RitaCucchiara,Member,IEEE

Abstract—Movingshadowsneedcarefulconsiderationinthedevelopmentofrobustdynamicsceneanalysissystems.Movingshadowdetectioniscriticalforaccurateobjectdetectioninvideostreamssinceshadowpointsareoften

misclassifiedasobjectpoints,causingerrorsinsegmentationandtracking.Manyalgorithmshavebeenproposedintheliteraturethatdealwithshadows.However,acomparativeevaluationoftheexistingapproachesisstilllacking.Inthispaper,wepresentacomprehensivesurveyofmovingshadowdetectionapproaches.Weorganizecontributionsreportedintheliteratureinfourclassestwoofthemarestatisticalandtwoaredeterministic.Wealsopresentacomparativeempiricalevaluationofrepresentativealgorithmsselectedfromthesefourclasses.Novelquantitative(detectionanddiscriminationrate)andqualitativemetrics(sceneandobjectindependence,flexibilitytoshadowsituations,androbustnesstonoise)areproposedtoevaluatetheseclassesofalgorithmsonabenchmarksuiteofindoorandoutdoorvideosequences.Thesevideosequencesandassociated“ground-truth”dataaremadeavailableathttp://cvrr.ucsd.edu/aton/shadowtoallowforothersinthecommunitytoexperimentwithnewalgorithmsandmetrics.IndexTerms—Shadowdetection,performanceevaluation,objectdetection,segmentation,trafficsceneanalysis,visualsurveillance.

æ

1

INTRODUCTION

DETECTIONandtrackingofmovingobjectsisatthecoreofmany

applicationsdealingwithimagesequences.Oneofthemainchallengesintheseapplicationsisidentifyingshadowswhichobjectscastandwhichmovealongwiththeminthescene.Shadowscauseseriousproblemswhilesegmentingandextractingmovingobjectsduetothemisclassificationofshadowpointsasforeground.Shadowscancauseobjectmerging,objectshapedistortion,andevenobjectlosses(duetotheshadowcastoveranotherobject).Thedifficultiesassociatedwithshadowdetectionarisesinceshadowsandobjectssharetwoimportantvisualfeatures.First,shadowpointsaredetectableasforegroundpointssincetheytypicallydiffersignificantlyfromthebackground.Second,shadowshavethesamemotionastheobjectscastingthem.Forthisreason,theshadowidentificationiscriticalbothforstillimagesandforimagesequences(video)andhasbecomeanactiveresearcharea,especiallyintherecentpast.Itshouldbenotedthat,whilethemainconceptsutilizedforshadowanalysisinstillandvideoimagesaresimilar,typically,thepurposebehindshadowextractionissomewhatdifferent.Inthecaseofstillimages,shadowsareoftenanalyzedandexploitedtoinfergeometricpropertiesoftheobjectscausingtheshadow(“shapefrom

.A.PratiandR.CucchiaraarewiththeDipartimentodiIngegneria

dell’Informazione,Universita

`diModenaeReggioEmilia,ViaVignolese,905/b,Modena,Italy.E-mail:{prati.andrea,cucchiara.rita}@unimore.it..I.MikiciswithQ3DMInc.,10110SorrentoValleyRoad,SuiteB,SanDiego,CA92121.E-mail:imikic@http://www.51wendang.com.

.M.M.TrivediiswiththeComputerVisionandRoboticsResearchLaboratory,DepartmentofElectricalandComputerEngineering,Uni-versityofCalifornia,SanDiego,9500GilmanDrive,LaJolla,CA92037.E-mail:trivedi@ece.ucsd.edu.Manuscriptreceived7June2001;revised19Aug.2002;accepted18Jan.2003.

RecommendedforacceptancebyP.Anandan.

Forinformationonobtainingreprintsofthisarticle,pleasesende-mailto:tpami@http://www.51wendang.com,andreferenceIEEECSLogNumber114319.

0162-8828/03/$17.00ß2003IEEE

PublishedbytheIEEEComputerSociety

shadow”approaches)aswellastoenhanceobjectlocalizationandmeasurements.Examplesofthiscanbefoundinaerialimageanalysisforrecognizingbuildings[1],[2],forobtaining3Dreconstructionofthescene[3],orevenfordetectingcloudsandtheirshadows[4].Anotherimportantapplicationdomainforshadowdetectioninstillimagesisforthe3Danalysisofobjectstoextractsurfaceorientations[5]andlightsourcedirection[6].

Shadowanalysis,consideredinthecontextofvideodata,istypicallyperformedforenhancingthequalityofsegmentationresultsinsteadofdeducingsomeimagingorobjectparameters.Intheliterature,shadowdetectionalgorithmsarenormallyassociatedwithtechniquesformovingobjectsegmentation.Inthispaper,wepresentacomprehensivesurveyofmovingshadowdetectionapproaches.Weorganizecontributionsreportedintheliteratureinfourclassesandpresentacomparativeempiricalevaluationofrepresentativealgorithmsselectedfromthesefourclasses.Thiscomparisontakesintoaccountboththeadvantagesandthedrawbacksofeachproposalandprovidesaquantitativeandqualitativeevaluationofthem.Novelquantitative(detectionanddiscriminationrate)andqualitativemetrics(sceneandobjectindependence,flexibilitytoshadowsituations,androbustnesstonoise)areproposedtoevaluatetheseclassesofalgorithmsonabenchmarksuiteofindoorandoutdoorvideosequences.Thesevideosequencesandassociated“ground-truth”dataaremadeavailableathttp://cvrr.ucsd.edu/aton/shadowtoallowforothersinthecommunitytoexperimentwithnewalgorithmsandmetrics.Thisavailabilityfollowstheideaofdata-sharingembodiedinCallforComparison,liketheprojectofEuropeanCOST211Group(seehttp://www.iva.cs.tut.fi/COST211/forfurtherdetails).

Inthenextsection,wedevelopatwolayertaxonomyforsurveyingvariousalgorithmspresentedintheliterature.Eachapproachclassisdetailedanddiscussedtoemphasizeitsstrengthsanditslimitations.InSection3,wedevelopasetofevaluationmetricstocomparetheshadowdetectionalgorithms.ThisisfollowedbySection4,wherewepresentaresultsofempiricalevaluationoffourselectedalgorithmsonasetoffivevideosequences.Thefinalsectionpresentsconcludingremarks.

2TAXONOMYOFSHADOWDETECTIONALGORITHMS

Mostoftheproposedapproachestakeintoaccounttheshadowmodeldescribedin[7].Toaccountfortheirdifferences,wehaveorganizedthevariousalgorithmsinatwo-layertaxonomy.Thefirstlayerclassificationconsiderswhetherthedecisionprocessintroducesandexploitsuncertainty.Deterministicapproachesuseanon/offdecisionprocess,whereasstatisticalapproachesuseprob-abilisticfunctionstodescribetheclassmembership.Introducinguncertaintytotheclassmembershipassignmentcanreducenoisesensitivity.Inthestatisticalmethods(see[8],[9],[10],[11],[12]),theparameterselectionisacriticalissue.Thus,wefurtherdividethestatisticalapproachesinparametricandnonparametricmethods.Thestudyreportedin[8]isanexampleoftheparametricapproach,whereas[10],[11]areexamplesofthenonparametricapproach.Thedeterministicclass(see[6],[7],[13],[14])canbefurthersubdivided.Subclassificationcanbebasedonwhethertheon/offdecisioncanbesupportedbymodel-basedknowledgeornot.Choosingamodel-basedapproachundoubtedlyachievesthebestresults,butis,mostofthetime,toocomplexandtimeconsumingcomparedtothenonmodel-based.Moreover,thenumberandthecomplexityofthemodelsincreaserapidlyiftheaimistodealwithcomplexandclutteredenvironmentswithdifferentlightingconditions,objectclasses,andperspectiveviews.

Itisalsoimportanttorecognizethetypesof“features”utilizedforshadowdetection.Basically,thesefeaturesareextractedfromthreedomains:spectral,spatial,andtemporal.Approachescanexploitdifferentlyspectralfeatures,i.e.,usinggraylevelorcolor

Word文档免费下载Word文档免费下载:2003_Detecting Shadows_Algorithms and Evaluation (共6页,当前第1页)

2003_Detecting Shadows_Algorithms and Evaluation相关文档

最新文档

返回顶部