AccuratePersonTrackingThroughChangingPosesfor
Multi-viewActionRecognition
PradeepNatarajan
RaytheonBBNTechnologies
10MoultonSt
CambridgeMA02138,USA
pradeepn@bbn.com
UniversityofSouthern
California
LosAngelesCA90089,USA
PrithvirajBanerjee
pbanerje@usc.edu
UniversityofSouthern
California
LosAngelesCA90089,USA
RamNevatia
nevatia@usc.edu
ABSTRACT
Actionsbyhumansinreal-worldsettingsinvolvelargechangesintheperson’sposeandtherelativeorientationwithrespecttothecamera.Persontrackingalgorithmsoftenfailundersuchconditions,sincetheyworkbydetectingandtrackingpeopleinafewknownposes(typicallystanding).Further,duetoocclusionsandsimilarityofclothingwithbackground,foregroundsilhouettesaretypicallyverynoisy.Wepresentanapproachwhichaddresstheseproblemsby rstaccu-ratelytrackingapersonthroughchangingposeandbrokenforegroundblobs.Duringthetrackingprocesswealsoes-timatetherelativeorientationandscaleoftheperson.Werepresenttheposeofthepersonineachtrackwindowus-ingagrid-of-centroidsmodel,andrecognizetheactionbymatchingwithasetofkeyposes,ineachframe.Wetestedourapproachinadatasetcollectedinarealgrocerystore,andreportbetterthan≈82.5%accuracyforframe-by-framerecognitionofactions.
1.INTRODUCTION
Systemsforautomaticanalysisandrecognitionofactionsfromvideoshaveseveralcompellingapplicationsincludinghuman-computerinteraction,visualsurveillance,searchandretrievalamongothers.Whilethegeneralproblemofactionrecognitioninunconstrainedvideosisextremelydi cult,simplemethodsthatworkinwell-constrainedenvironmentscanhaveasigni cantpracticalimpact.Withthisinmind,wefocusontheproblemoftrackingandrecognizingsinglepersonactions,inlargelystaticbackground,indoorenviron-ments.Suchconditionsarecommoninlibrariesandsmallgrocerystoreswheremostaislesaresparselypopulated.Extensiveresearchhasbeenconductedinactionrecog-nition,duetotherangeofpotentialapplications.How-ever,theapplicabilityofcurrentmethodsremainlimited,sincelow-leveltrackingandposeestimationremainsex-tremelyhard.Whilesigni cantprogresshasbeenmadeCorrespondingauthor
Permissiontomakedigitalorhardcopiesofallorpartofthisworkforpersonalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforpro torcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationonthe rstpage.Tocopyotherwise,torepublish,topostonserversortoredistributetolists,requirespriorspeci cpermissionand/orafee.
ICVGIP’10,December12-15,2010,Chennai,India
Copyright2010ACM978-1-4503-0060-5/10/12...$10.00.
155
inpedestriandetection[5][28]andtracking[15],suchdetect-and-associatemethodscanhandleonlyafewknownposes(typicallyupright).Thisisbecausethedetectorsneedalargesetoftrainingimagesforeachposeofinterest,andhencearenotscalableforactionsinvolvingextensiveposearticulation.Thismakesitnecessarytodevelopsuitabletechniquestopredictandinterpolatebetweentwodetec-tionsofknownposes,whichmustalsoberobusttofalsealarmsandmisseddetections.
Further,manyactionrecognitionapproachesrelyonex-tracting2Dfeatureslikespatio-temporalinterestpoints[13]andlearningsuitableclassi ers.Thelearnedmodelsaretypicallynotgeneralizableacrossdatasets,viewpointsandscale.Thiswouldeitherrequirecollectingaprohibitivelylargetrainingset,orretrainingthemodelsforeachnewdataset.Toaddressviewpointvariations,[17][11][20]use3DMotionCapture(Mocap)dataofactionsrenderedfrommultipleviewpoints,whicharethenmatchedto2Dfeatures.However,suchMocapdataistimeconsumingtocollectandrequiresexpensiveequipment.Alsorenderingthe3Dmod-elsfrommultipleviewpointstakesalongtime(3-4days),andrequiresproprietarysoftwarelikePoser.
Weaddresstheselimitationsby rstpresentinganalgo-rithmtotrackasinglepersonthroughlargechangesinori-entation,scaleandposearticulation.Duringtrackingwealsoestimatetheapproximateorientationandheightofthepersonwithrespecttothecamera.Werepresenttheposeineachtrackwindowofeachframe,usingagrid-of-centroidsfeaturevectorthatisrobusttolargeerrorsinbackgroundsubtraction.Wenextrecognizeactionsbymatchingthefea-turevectorwiththeactionmodels,whichisrepresentedus-ingasequenceofkeyposesandmotionmodels.Thekeyposesarerepresentedusingstick gures,whosearticulationscanbespeci edmanually.Further,welearnourmotionmod-elsformulti-viewactionrecognitionfromafewvideosam-ples,withoutrequiringcameracalibration.Ourapproachpresentsane ectiveandeasy-to-useframeworkforframe-by-frameactionrecognitionincommonindoorsettings.Wedemonstrateourmethodonadatasetcollectedinarealgro-cerystorewith8di erentactorsand3actions,undervary-ingclothing,lightingandviewpoints.Oursystemachieves≈82.5%accuracyandrunsat5fpsonastandarddesktopcomputer.
Intherestofthepaper,wewilldiscussrelatedworkinsection2,overviewofourapproachinsection3,trackingalgorithminsection4,methodsforlearningkeyposesandmotionmodelsinsection5,matchingandrecognitioninsec-tion6andresultsinsection7.
热门文档