人体运动分析A new approach for classification of human g

Anewapproachforclassificationofhumangaitbasedontime-frequencyfeaturerepresentationsIrenaOrovic´a,n,SrdjanStankovic´a,MoenessAminbaUniversityofMontenegro,FacultyofElectricalEngineering,81000Podgorica,MontenegrobCenterforAdvancedCommunications,VillanovaUniversity,Villanova,PA19085,USAarticleinfoArticlehistory:Received21January2010Receivedinrevisedform20August2010Accepted27August2010Availableonline8September2010Keywords:Time-frequencydistributionsRadarsignalsHumanwalkingMicro-DopplersignaturesabstractWeintroduceanewandsimpletechniqueforhumangaitclassificationbasedonthetime-frequencyanalysisofradardata.Thefocusisontheclassificationofarmmovementstodiscernfreevs.confinedarmswingingmotion.Thelattermayariseinhostagesituationormaybeindicativetocarryingobjectswithoneorbothhands.Themotionsignaturescorrespondingtothearmandlegmovementsarebothextractedfromthetime-frequencyrepresentationofthemicro-Doppler.Thetime-frequencyanalysisisperformedusingthemultiwindowS-method.WiththeHermitefunctionsactingasmultiwindows,itisshownthattheHermiteS-methodprovidesanefficientrepresentationofthecomplexDopplerassociatedwithhumanwalking.TheproposedhumangaitclassificationtechniqueutilizesthearmpositiveandnegativeDopplerfrequenciesandtheirrelativetimeofoccurrence.Itistestedonvariousrealradarsignalsandshowntoprovideanaccurateclassification.&2010ElsevierB.V.Allrightsreserved.1.IntroductionAmongseveralpossibletechnologies,includingacous-tics,thermal,optical,andradiofrequency,RF-basedtechnologyisconsideredanattractivemodalityforhumanmotiondetectionandclassifications,asitcanbeappliedunderallweather,light,smoke,andfortargetsobstructedbyopaquematerial.Recently,radarhasbeensuccessfullyusedinurbansensingapplicationsandthroughwallimaging[1–5].Thispaperconsidersradarsfortheclassificationofhumangaitbasedondistinctionsinthewalkingperson’sarmmovements.Inparticularthreetypesofwalkingmotionsareofinterest:(1)Freearm-motion(FAM)characterizedbyswingingofbotharms,(2)Partialarm-motion(PAM)whichcorrespondstoamotionofonlyonearm,and(3)Noarm-motion(NAM)whichcorrespondstonomotionofeitherarm.TheNAMisreferredtoasastrollerorsaunterer[6].Thelasttwoclassesarecommonlyassociatedwithapersonwalkingwithhis/herhand(s)inthetrouserpocketsorapersoncarryinglightsmallorlargeheavyobjects,respectively.Allthreecategoriesareconsideredimportantinlawenforcementandhomelandsecurityoperations.Theradarmicro-Dopplerforhumangaithasbeenanactiveareaofresearchforthelastdecade[7,8].InadditiontothemainDopplershiftduetothemotionofthehumantorso,therelativemotionsofthelimbstothebodyintroducemicro-Dopplerwhichpresentsitselfasatime-varyingfrequencyshift.ThecomplexnonstationaryDopplersignatureofhumanwalkingcanberevealedviaajointtime-frequencysignalrepresentationinlieuofthetraditionalFouriertransform,oftheradarreturn.Thedegreeofclarityanddepictionofthetime-dependentDopplerfrequencyforeachpartofthehumanbodyinmotioncanvarydependingonthetime-frequencyanalysistoolemployed.Comparedtoothermethods,time-frequencydistributions,whichcapturetheinstanta-neousfrequencylawsaremostsuitablefortheunderlyingapplication[9–13].ContentslistsavailableatScienceDirectjournalhomepage:www.elsevier.com/locate/sigproSignalProcessing0165-1684/$-seefrontmatter&2010ElsevierB.V.Allrightsreserved.doi:10.1016/j.sigpro.2010.08.013nCorrespondingauthor.Tel.:[phone];fax:[phone].E-mailaddress:[e-mail](I.Orovic´).SignalProcessing91(2011)1448–1456ClassificationsoftheabovetypesofhumangaitwereconsideredinRefs.[6],[14–16].TheworkinRef.[6]onlydealtwithFAMandNAMtypesandusedSpectrogramforthedistributionofDopplersignalpowerinthetime-frequencydomain.Thiswork,thoughimportant,didnotconsiderdistinctionsinthetypesofmotion,butratherestimatedthehumanwalkingparametersbyminimizingthedifferencebetweensimulatedThalmannmodel[16]andrealmeasurements.HumangaitclassificationsofthethreetypesFAM,PAM,andNAM,werediscussedinRef.[14]basedonsubspacelearningusingprincipalcompo-nentanalysis(PCA).Thetrainingsetconsistsoffeaturevectorsdefinedaseithertimeorfrequencysnapshotstakenfromthespectrogramofradarbackscatter.Thismethod,althoughgeneratedpromisingclassificationresults,isnonparametricanddidnotexplicitlyutilizetheperiodicandevolvingnatureofthehumangaitinthethreemotiontypes.Thetime-frequencyclassifierinRef.[15]applieddistancemeasuresbetweentrainingandtestsetsrepresentedbythetime-frequencydistributionsofthecorrespondingDopplersignals.Thisclassifierisalsoanonparametricmethod.ItneitherselectsnordoesitseparatethekeyanddistinctiveDopplerfeatures,asso-ciatedwiththearms’motions.TheclassifieremployedinRef.[17]wasbasedonSVMandconsideredseveraltypesofhumanmotions,includingrunningandcrawling.Itisaparametrictechniqueanduseddifferentnumericalfea-turesofthehumanmotion.Thistechnique,however,didnotconsidertheclassificationoftheabovethreeclassesandtheircorrespondingsalientfeature.Ourcontributiontotheabovegaitclassificationproblemistwofold.WeapplytherecentlyintroducedmultiwindowS-method[18]asahightime-frequencyconcentrationtechnique,inlieuofWignerdistribution,spectrograms,orothertime-frequencysignalr