基于时间卷积轻量级网络的人体运动预测（英文）

基于时间卷积轻量级网络的人体运动预测（英文）Abstract:Body motion prediction is a critical problem in many applications, su

基于时间卷积轻量级网络的人体运动预测（英文） Abstract: Bodymotionpredictionisacriticalprobleminmany applications,suchashuman-computerinteraction,virtualreality, andsportsanalysis.Althoughconventionaldeeplearningmodels canachievehighaccuracyinmotionforecasting,theyaretypically computationallyexpensiveandrequiremassiveamountsof trainingdata.Inthiswork,weproposeatime-convolutional lightweightneuralnetwork(TCLNet)forhumanactionprediction, whichcombinesdepth-wiseseparableanddilatedconvolutions withamulti-scaleresiduallearningscheme.Ourapproach achievesstate-of-the-artpredictionaccuracywithfewer parametersandlowercomputationalcost. Introduction: Overthepastdecade,therehasbeenagrowinginterestin predictinghumanbodymotionduetoitssignificanceinvarious applications.Forexample,inhuman-computerinteraction,motion predictioncanenhanceimmersiveexperiencesbyenablingnatural interactionsbetweenhumansandcomputers.Invirtualreality applications,motionpredictioncanimprovetherealismofavatar movements.Insportsanalysis,motionpredictioncanassist coachesinevaluatingathlete'sperformanceanddeveloping trainingstrategiesbasedonthepredictedoutcomes. However,predictinghumanbodymotionisachallengingtask duetothecomplexityofhumanmovementsandtheneedfor accuratetiming.Also,deeplearningapproachestypicallyrequire large-scaletrainingdataandhighcomputationalresources,which canbeprohibitiveinsomepracticalusecases. Toaddressthesechallenges,weproposeanovel time-convolutionallightweightneuralnetwork(TCLNet)for humanactionprediction.Ourmethodtakesasinputasequence ofhumanjointcoordinatesandpredictsthefuturebody movements.TCLNetcombinesdepth-wiseseparableanddilated convolutionswithamulti-scaleresiduallearningschemeto achievehighaccuracyinmotionforecastingwithfewer parametersandlowercomputationalcost. RelatedWork: