This paper presents our work on learning high level structure from human motion sequences, and its applications in human figure tracking. We use a structured representation (“primitives ” and their transitions) of complex motion and propose a two-step un
ACCV2002:The5thAsianConferenceonComputerVision,23–25January2002,Melbourne,Australia.
LearningAHighlyStructuredMotionModelfor3DHumanTrackingTaoZhao
UniversityofSouthernCalifornia
LosAngeles,CA90089
TianshuWang
Arti cialIntelligenceandRoboticsLab
Xi’anJiaotongUniversity
Heung-YeungShum
MicrosoftResearch,ChinaBeijing10080,China
taozhao@iris.usc.edu
Abstract
tswang@asia.comhshum@microsoft.com
Thispaperpresentsourworkonlearninghighlevelstruc-turefromhumanmotionsequences,anditsapplicationsinhuman guretracking.Weuseastructuredrepresentation(“primitives”andtheirtransitions)ofcomplexmotionandproposeatwo-stepunsupervisedlearningapproachtore-coverthenatural“primitives”fromunsegmented3D-motioncapturedsequencesofcomplexhumanmotion.Thestruc-turerecoveryisdoneundertheMDL(minimumdescriptionlength)paradigm.ThenthelearntdynamicmodelofhumanmotionisusedintheCONDENSATIONframeworktosuc-cessfullytrackhumanmotioninavideosequence.Experi-mentalresultsofballetdancingsequencesdemonstratethatourapproachworkswell.Thelearntstructureisalsousedtosynthesizenewvideosequences.
1Introduction
Humanmotiontrackingandunderstandinghavebeenac-tiveresearchtopicsincomputervisionforalongtimebe-causeoftheirwideapplicationsinrecognition,animationandhuman-computerinteraction.Goodsurveyonvisualanaly-sisofhumanmotionandvision-basedmotion-capturecanbefoundin[7],[13].However,3Dhumantrackingfroma2Dvideosequenceremainslargelyunsolved.
In3Dhumantracking,akinematicsmodelisgenerallyusedandtheproblemistorecoverthejointanglesfromtheimagesequence.Manyfactorscontributetothedif cultyof3Dhuman guretracking:
kinematicshassingularposes;
cameraprojectionlosesdepth,andintroducesanobser-vationsingularityandre ectivesymmetry;largenumberofDOFs(20+forfullbody);complexhumandynamics;
variousimagenoisesincludingcomplexbackground,non-rigidmotionofbodyandclothing,changeofillu-mination,etc.
Thisworkwasdonewhilethe rsttwoauthorswerevisitingMicrosoftResearch,China.
Someofthedif cultiescanbealleviatedbyintroducinganappropriatedynamicmodel(i.e.,howtheobjectmoves)whichservesasaprior.Forcomplexhumanmotion,suchasdancing,gymnasticsandKungFu,whicharemadeupofanumberofbasicmovesor“buildingblocks”,asingledynamicmodelisingeneralinsuf cient,instead,astructureddynamicmodelispreferred.
Muchpreviousworkhasbeendoneinmodelingcomplexhumanmotionmodelandtheycanbelargelycategoriedintotwoclasses.The rstclassisbysupervisedlearning.Mixturemotionmodelisusedfortrackingin[9].Buttheprimitivesarepre-de nedandsegmentedmanuallyfortraining.
Thesecondclassofapproach,unsupervisedorsemi-unsupervisedhumanmotionmodeling,avoidssuchtediousanderrorproneprocessofmanualsegmentation.In[3],HMM(hiddenMarkovmodel)islearntforhumanlocomotion(walking,running).ButthetopologyoftheHMMisgivenanditisdif culttoextendittomorecomplexmotion.In[1],eachprimitivefollowsadifferentdynamiclaw(acceler-ation)whichcanbeusedtodifferentiateeachother.VariablelengthMarkovmodels(VLMM)[6]werelearnttomodelhu-manbehavior.However,simpleheuristicssuchaslowveloc-itypointsattheboundaryoftwoprimitiveswasemployedforsegmentation.SLDS(switchinglineardynamicsystems)arelearntin[14]forclassifyinghumanmotion.[2]alsolearnedHMMsforcomplexmotionsuchasdancing.ItemployedaverycomplicatedoptimizationprocessandthehighlevelstructureisstillhiddenintheHMM.
Inthispaper,weadopttheunsupervisedapproachandaimtorecovertheoriginalprimitivesinasystemeticway.Mini-mumdescriptionlength(MDL)criterionisusedforstructurerecoveryandanexplicittwo-stepcoarseto nemethodispro-posed.Inbothsteps,MDLguaranteestheconcisenessoftherecovery,thusleadingtooriginalornearlyoriginalprimitives.Oncelearnt,thestructuredrepresentationcanthenbeusedintrackingandsynthesis.AnoverviewofoursystemisshowninFig.1.
Itisworthmentioningtheballetdancingdataweuseinourexperiments.Tolearnthemotionmodel,weuse10sequencesofcaptured3Dmotionofaballettutorial,4428frames(about3min)intotal.Wetrytomodelthearmmovements(4DOFs
Learning a highly stru... 6页 免费 Bayesian ... Learning to Track 3D Human Motion from ...An important sub-case is model-based tracking, ...
Learning to reconstruct ... 暂无评价 28页 免费 Learning a highly struct....(2003). Kinematic Jump Processes For Monocular 3D Human Tracking. Int. ...
Impact of Dynamic Model Learning on Classi?...tracking of human motion and contrast them to ...We develop a novel approximate structured ...
that combines Gaussian process learning, a particle...lter for 3D human pose tracking. Our method can...“Twin Gaussian processes for structured prediction,...
3D Modeling for Capturing Human Motion from ...demand highly realistic Human Body Models (HBMs)....frames or with a learning-based approach [9]. ...
from images, which is a highly non-linear and ...structured prediction on computer vision tasks, ...culty learning an implicit model of the ...
tracking of human motion (see [1, 9] for a ...lters [17, 4] or structured variational inference...Learning and recognizing human dynamics in video ...
The basic idea is to start human tracking ...motion patterns in this space are structured in ...A common problem with learning-based approaches is...
Learning and tracking ... 暂无评价 7页 免费 ...Motion Capture, Articulated Human Tracking, 3D ...motion parameters are then more highly constrained....
We propose a learning approach to model non-...motion tracking, including the high dimensionality ...The human body is highly constrained by strict ...
热门文档