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

2.PREVIOUSRESEARCH3.THEPROPOSEDMCSHMM

Thisresearchfocusesonexploringtheinteractionbetweenmultiplesemanticstructureswiththeaimtoimprovetheoverallminingperformance.Speci cally,itwasobservedthatcameraviewsandplaytypesarequiterelatedtoeachotherduringagame.Forexample,afterashotofthecentralviewaccompaniedbyashortplay,thecameraviewinthenextshotislikelytoremaininthesameview.ItistemptingtoinvoketheCoupledHMM(CHMM)proposedin[1]forourproblem.However,theCHMMusuallyassumesstronginteractionbetweentwoMarkovchainsthatmayoverweightorevencorrupttheMarkovianpropertywithineachchainiftheassumptionisnottrue,asobservedinourpractice.Therefore,weadvanceamulti-channelSHMMmodel(MC-SHMM)thatinvolvestwoparallelSHMMsandatwo-layerhierarchicalMarkovianstructure,asshowninFig.3.

Multi-channel Segmental Hidden Markov Models for

Sportvideosnormallyhavetwofundamentalsemanticstructuresregarding“what”and“where”.OrinthecaseofAmericanfootballvideos,theyarecalled“playtypes”and“cameraviews”[4,5],whichcouldbeusedasbuildingblockstodeliverhigherlevelsemantics.Eachsematicstructurein-volvesmultiplesemanticeventsthatareassumedtofollowa rst-orderMarkovchainovertime,asshowninFig.1.

Multi-channel Segmental Hidden Markov Models for

Multi-channel Segmental Hidden Markov Models for

Thetwosemanticfootballvideos,cameraviews(left)andplaytypes(right).

2.1ProblemFormulation

GivenavideosequencehasbeenpartitionedintoTshots

ofvariable-length,eachofwhichcorrespondstoaspeci cplayandacameraview,thevideominingtaskherecanbeformulatedastheinferenceproblemofagenerativemodelwherethelatentstateindicatesthesemanticeventofashot,andtheobservationmodelcharacterizesvisualfeaturesex-tractedfromallframesinashotwithrespecttodi erentse-manticevents.Thefeaturesusedforviewanalysisincludesthecolordistributionandtheangleofyardlines[4],andthatofplayanalysisinvolvesmotionvectorsextractedbyane -cientoptical owtechnique[12].How

Multi-channel Segmental Hidden Markov Models for

tocapturerichstatis-ticsofframe-wiseobservationstosupportshot-wisestatesisakeyissueforourpreviousresearch.Atwo-layergen-erativemodelwasproposedin[5]thatde nesstatesonaGaussianmixturemodel(GMM)(notontheobservation),andinvolvesanembeddedHMM-GMMstructure.

Figure3:IntheMCSHMM,the rstlayerincludestwo

SHMMs,andthesecondlayercapturestheinteraction

(j,i)

oftwochannelsatthe rstlayer.Oji={O1:τi}representsallframe-wiseobservationsofshotiinchannelj∈{1,2}.

3.1ModelParameters

TheuniquefeatureoftheMCSHMMistheintegrationofbothparallelandhierarchicalstructuresonlatentstatesandtheincorporationofasegmentalobservationmodel,whichallowtheMCSHMMhavemore exibility,capacity,andfunctionalitythantheCHMM,SHMMandHHMM.Intheviewofgenerativemodels,boththedynamicmodel(amonghiddenstates)andtheobservationmodelintheMCSHMMhaveatwo-layeredstructurethatgreatlyen-hanceitscapabilityoflearningandinference.Speci cally,

j

atthe rst-layer,S={Si|i=1,...,T;j=1,2}denotes

j

thestatesequenceoftwochannelswhereSidenotesthe

(j,i)j

stateofshotiinchannelj,andOi={O1:τi|j∈(1,2)}indicatesobservationsofshotiwithτiframesinchannel

12

j.F={Fi=(Si,Si)|i=1,...,T}representsthestatese-quenceatthesecondlayerwhereeachstateconsistsoftwostatesatthelowerlayer.Therefore,theMCSHMM’sparam-etersetΘincludesfollowingcomponentsΘ={Π,A, }: Initialprobabilities:

1212

Π={P(S1),P(S1),P(F1|S1,S1)};

aretheobservationsofshotiwithτiframes.

iFigure2:ThesegmentalHMM(SHMM)whereOj=1:τi

2.2SegmentalHiddenMarkovModels

TheSHMMwasoriginallyproposedforspeechsignalpro-cessing[6]thatinvolvesasegmentalmodeltorepresenta

variable-lengthsequenceofobservations,asshowninFig.2.IntheSHMM,eachstateemitsasequenceofobservationsthatiscalledasegment.Ineachsegment,allobservationareconditionallyindependentgiventhemeanofthesegment,leadingtoaclosed-formlikelihoodfunctionfortheobserva-tionmodel.In[4],theSHMMwasshowne ectiveonview-basedvideomingcomparedwiththetraditionalHMMsandtheembeddedHMM-GMMin[5].HencetheSHMMwillserveasthebuildingblockforourproposedmodel.

(1)

Transitionprobabilities:

A={Aw,w=1,2,3},

where

A1A2A3

11

={P(Si=m|Si 1=n,Fi 1=l)},22={P(Si=m|Si 1=n,Fi 1=l)},12={P(Fi=l|Si=m,Si=n)};

(2)

上一页第2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