资源描述：
《Online Learning of Robust Facial Feature Trackers》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、OnlineLearningofRobustFacialFeatureTrackersTimSheerman-Chase,Eng-JonOngandRichardBowdenCVSSP,UniversityofSurrey,Guildford,SurreyGU27XH,UnitedKingdomt.sheerman-chase,e.ong,r.bowden@surrey.ac.ukAbstractface[5,13,9,14],butthesemodelsonlyapproximatetheshapeofthefa
2、ceatnearfrontalheadpose.Alternatively,Thispaperpresentsaheadposeandfacialfeaturees-various3Dheadmodelscanbeused[7,1,12,4,2].Thesetimationtechniquethatworksoverawiderangeofposemorecomplexmodelsrequireaccurateinitializationandarevariationswithoutaprioriknowledge
3、oftheappearancecomputationallyexpensive.3Dheadmodelscanrangefromoftheface.UsingsimpleLKtrackers,headposeisesti-simplecylindricalandellipsoidalmodelstocomplexpolyg-matedbyLevenberg-Marquardt(LM)poseestimationus-onalapproximation,usuallyusedforposeestimation.Man
4、yingthefeaturetrackingasconstraints.Factoredsamplingofthesetechniquesusetemplateupdate,byincrementallyandRANSACareemployedtobothprovidearobustposemodifyingtheexpectedappearancetocorrespondwiththeestimateandidentifytrackerdriftbyconstrainingoutliersobservedappe
5、aranceandtoachieveposeinvarianttrack-intheestimationprocess.Thesystemprovidesbothaheading.Moreaccuratemodel-basedfacialfeaturetrackingcanposeestimateandthepositionoffacialfeaturesandisca-beachievedusingarealistic3Dfacemodelofsufficientde-pableoftrackingoverawid
6、erangeofheadposes.tailforrendering.Thisallowsforarealisticreconstructionofthevisualappearanceoftheface[11].However,thesemodelsarecomplicatedanddeformingthemodeltofitthe1.Introductionimageisusuallynon-trivial.ThispaperpresentsanapproachforthetrackingoffacialThen
7、extclassofapproachescouplesasetofpose-featuresandposeestimateoftheheadthroughoutavideospecifictrackerswithaswitchingmechanism(e.g.posees-sequencewithoutanapriorimodelofappearance.Theap-timator)todecidewhichofthesetrackerstouse.Anexam-proachusesonlinelearningtob
8、uildamodelofappearancepleisproposedbyKanaujiaetal.2006[6]wheremultipleon-the-flyusingageneric3Dshapemodeltoremovetrack-2Dpose-specificActiveShapeModels(ASMs)arecoupledingdriftinheren
