Quantifying and recognizing human movement pattern

QuantifyingandRecognizingHumanMovementPatternsfromMonocularVideoImages-PartI:ANewFrameworkforModelingHumanMotionR.D.GreenandL.GuanAbstract–Researchintotrackingandrecognizinghumanmovementhassofarbeenmostlylimitedtogaitorfrontalposing.PartIofthispaperpresentsaContinuousHumanMovementRecognition(CHMR)frameworkwhichformsabasisforthegeneralbiometricanalysisofcontinuoushumanmotionasdemonstratedthroughtrackingandrecognitionofhundredsofskillsfromgaittotwistingsaltos.PartIIofthispaperpresentsCHMRapplicationstothebiometricauthenticationofgait,anthropometricdata,humanactivitiesandmovementdisorders.InPartIofthispaper,anovel3Dcolorclone-body-modelisdynamicallysizedandtexturemappedtoeachpersonformorerobusttrackingofbothedgesandtexturedregions.TrackingisfurtherstabilizedbyestimatingthejointanglesforthenextframeusingaforwardsmoothingParticlefilterwiththesearchspaceoptimizedbyutilizingfeedbackfromtheCHMRsystem.Anewparadigmdefinesanalphabetofdynemes,unitsoffull-bodymovementskills,toenablerecognitionofdiverseskills.UsingmultipleHiddenMarkovModels,theCHMRsystemattemptstoinferthehumanmovementskillthatcouldhaveproducedtheobservedsequenceofdynemes.Thenovelclone-body-modelanddynemeparadigmpresentedinthispaperenabletheCHMRsystemtotrackandrecognizehundredsoffull-bodymovementskillsthuslayingthebasisforeffectivebiometricauthenticationassociatedwithfull-bodymotionandbodyproportions.R.D.GreeniswiththeHumanInterfaceTechnologyLab,UniversityofCanterbury,Christchurch,NewZealand.HewaswiththeSchoolofElectricalandInformationEngineering,TheUniversityofSydney,NSW2006,Australia,(e-mail:[e-mail]).Prof.L.GuaniswiththeDepartmentofElectricalandComputerEngineering,RyersonUniversity,Toronto,ONM5B2K3,Canada(correspondingauthor:phone:[phone]ext.6072;fax:[phone];e-mail:(e-mail:[e-mail]).1.IntroductionBiometricauthenticationdependsonsignificantmeasurablediversityofaparticularphysicalcharacteristic,suchasiris,fingerprint,signatureorgait.Themoredimensionsandlargerbetween-personvariabilityforeachdimension,thebetterthebiometric.Thegoalistoresolvetheapparentconflictofenhancingbetweenindividualvariationswhileminimizingwithinindividualvariations.With249degreesoffreedomandgooddiscriminationentropy,theirisbiometriciswellaheadofothersbyreliablyrecognizing9millionwithnofalsepositives[6]andwithprojectionsto1in10billion–morethanthepopulationofthisplanet.Otherbiometricssuchasfaceandgaitareordersofmagnitudeawayfromirisrecognitionaccuracyandunlikeiris,gaitandfaceareaffectedbyage,clothes,andaccessories,leavingmanyproblemsyettobesolved.However,biometricsengagingthewholebody,suchasgait,haveaplaceforlessproximalbiometricauthenticationwhereidentificationispossiblewithoutanyawarenessofthesubjecttominimizeriskofanidentitybeing‘faked’.PartIofthispaperpresentsaContinuousHumanMovementRecognition(CHMR)framework(Figure1)whichformsabasisforfull-bodybiometricanalysisofcontinuoushumanmotionandanthropometricdata.PartIIofthispaperpresentsCHMRapplicationstothebiometricauthenticationofgait,anthropometricdata,humanactivitiesandmovementdisorders.Figure1.Overviewofthecontinuoushumanmovementrecognitionframework.jjooiinnttaanngglleessvviiddeeooTTRRAACCKKIINNGGSSppaattiiaallSSeeggmmeennttaattiioonnParticleFilterRREECCOOGGNNIITTIIOONNTTeemmppoorraallSSeeggmmeennttaattiioonnHiddenMarkovModelssskkiillllnnaammeessHumanmovementiscommercially1trackedbyrequiringsubjectstowearjointmarkers/identifiers,anapproachwithhasthedisadvantageofsignificantsetuptime.Suchaninvasiveapproachtotrackinghasbarelychangedsinceitwasdevelopedinthe1970s.Usingalessinvasiveapproachfreeofmarkers,computervisionresearchintotrackingandrecognizingfull-bodyhumanmotionhassofarbeenmainlylimitedtogaitorfrontalposing[27].Variousapproachesfortrackingthewholebodyhavebeenproposedintheimageprocessingliteratureusingavarietyof2Dand3DshapemodelsandimagemodelsaslistedinTable1.AuthorsShapemodelImagemodelHoggCylindersEdgeRohrCylindersEdgeGavrila&DavisSuperquadricsEdgeDrummond&CipollaConicsEdgeGoncalvesetal.ConesEdgeKakadiaris&MetaxasDeformableEdgeWren&Pentland2DcolorblobsSkincolorblobsJuetal.Patches(2D)FlowBregler&MalikCylindersFlowWangetal.CylindersFlowCham&RehgPatches(2D)TemplateWachter&NagelConesFlow+EdgePlänkers&FuaDeformableSilhouette+DisparityDeutscheretal.ConesEdge+SilhouetteBrandOutlineSilhouettemomentsRosales&SclaroffOutlineSilhouettemomentsLiebowitz&CarlssonOutlineHand-markedjointsTaylorOutlineHand-markedjointsLeventon&FreemanOutlineHand-markedjointsTable1Comparisonofdifferenthumanbodymodels.1Commerciallyavailabletrackersarelistedat