文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 工程科技 > 城乡/园林规划 > Multi-channel Segmental Hidden Markov Models for
侵权投诉

Multi-channel Segmental Hidden Markov Models for

多通道分割

Multi-channelSegmentalHiddenMarkovModelsfor

SportsVideoMining

SchoolofElectricalandComputerEngineeringOklahomaStateUniversity,Stillwater,OK74078

YiDingandGuoliangFan

{yi.ding,guoliang.fan}@okstate.edu

ABSTRACT

Wepresentanewmulti-channelsegmentalhiddenMarkovmodel(MCSHMM)forsportsvideominingthatisauniqueprobabilisticgraphicalmodelwithtwoadvantages.Oneistheintegrationofbothhierarchicalandparallelstructuresthato ermore exibilityandcapacityofcapturingthein-teractionbetweenmultipleMarkovchains.TheotheristheincorporationofthesegmentalHMMthatrepresentsavariable-lengthsequenceofobservations.Especially,wede-velopamaximumaposteriori(MAP)estimatortooptimizemodelstructuresandmodelparameterssimultaneously.TheproposedMCSHMMisusedforAmericanfootballvideoanalysis,wheretwosemanticsstructures,playtypesandcameraviews,areinvolved.TheexperimentshowsthattheMCSHMMoutperformspreviousHMM-basedapproaches.

(whathappened?)andcameraviews(whereithappened?).Speci cally,weareinterestedindevelopingpowerfulgenera-tivemodelsthatareabletocapturetheinteractionbetweenmultiplesemanticstructuresandimprovetheoverallper-formance.Hereweproposeanewmulti-channelsegmentalhiddenMarkovmodel(MCSHMM)thatisinspiredbyrecentprogressongraphicalmodeltheory.Asacasestudy,wetestthenewMCSHMMforAmericanfootballvideoanalysis.

1.2Relatedworks

Therearetwomajormethodsinvideominingresearch[8].Oneisthediscriminativeapproaches,e.g.,[9],andtheotheristhegenerativeapproaches,e.g.,[11].Agenerativeap-proachisusuallypreferredwhenvideominingisperformedonalargedatasetduetoitsbettergenerality.TheDynamicBayesNetwork(DBN)[10]providesauni edprobabilisticframeworktorepresentvariousgenerativemodelswithdif-ferentgraphicalstructures.Recently,severalspeci cDBNs,i.e.,theFactorialHMM[7],theCoupledHMM(CHMM)[1],andthein uencemodel[14],wereproposedtoexplorecom-plexsemanticstructuresunderlyingmultipledataorfeaturestreams.Additionally,itisimperativetohavesimultaneousstructurelearningandparameterlearningforcomplexDBNstoensuretheire ectivenessande ciencyinpractice.In[2],theconceptsofentropicpriorandparameterextinctionwereproposedtosimplifyandoptimizethemodelstructurebyastatetrimmingprocess.In[13],thehierarchicalHMMs(HHMMs)wereusedforunsupervisedsportsvideomining,wherethereverse-jumpMarkovChainMonteCarlo(RJM-CMC)methodisembeddedinthelearningofHHMMstosupportmodelselectionandfeatureselection.

CategoriesandSubjectDescriptors

I.4.8[ImageProcessingandComputerVision]:SceneAnalysis

GeneralTerms

Algorithms

Keywords

sportsvideomining,entropicprior,hiddenMarkovmodels

1.INTRODUCTION

1.1Motivation

Videominingistodiscoverknowledge,patternsandevents,orcalledsemanticstructures,inthevideodata,anditsben-e tsrangefrome cientsummarizingofvideocontentofinteresttofacilitatingvideoaccessandretrievalinlargedatabases.Inthiswork,westudysportsvideominingwhererelativelyde nitesemanticstructuresexistandthathastremendouscommercialvalue.However,multipleseman-ticstructurescouldconcurinasportsvideo,e.g.,playtypes

1.3Ourworkandcontribution

Thisworkisbuiltonourpreviousresearch[3,5,4],wheretwosemanticstructureswereconsideredindependently,i.e.,playtypesandcameraviews.WeimprovedexistingHMMsbydevelopingmoree ectiveobservationmodelstocapturerichstatisticsofavariable-lengthsequenceofobservations.Thisworkisaimedathowtoexploretheinteractionbe-tweentwoMarkovchains(i.e.,semanticstructures).Therearetwocontributions.Oneistheintegrationofbothhier-archicalandparallelstructuresandtheincorporationofasegmentalobservationmodel.Theotheristheapplicationofentropicpriorandparameterextinctioninthemodeltrain-ingtooptimizethemodelstructureandmodelparameterssimultaneously.TheMCSHMMo ersmore exibility,func-tionality,andcapacitycomparedwithexistingHMMs,andholdsgreatpotentialforothervideominingapplications.

Permissiontomakedigitalorhardcopiesofallorpartofthisworkforpersonalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforpro torcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationonthe rstpage.Tocopyotherwise,torepublish,topostonserversortoredistributetolists,requirespriorspeci cpermissionand/orafee.

MM’08,October26–31,2008,Vancouver,BritishColumbia,Canada.Copyright2008ACM978-1-60558-303-7/08/10...$5.00.

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email