欢迎来到天天文库
浏览记录
ID:41230036
大小:216.00 KB
页数:12页
时间:2019-08-19
《HumanPosedetection:人体姿态检测》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、HumanPosedetectionAbhinavGolasS.ArunNairOverviewProblemPrevioussolutionsSolution,detailsProblemSegmentationofhumansfromvideocapturePosedetection(byfittingontobodymodel)Resistanttonoise(backgroundetc.)PreviousproceduresViewproblemassequentialprocessSegmentationPosedetect
2、ionProblems:Notusingpriorknowledgeof“whatahumanlookslike”insegmentationUsesonlyinformationfromdetected“foreground”forposedetectionAllavailableinformationnotusedSolutionCombinesegmentationandposedetectionasasinglestepUsesallavailableinformationinframe(forposedetection)Us
3、espriorknowledgeofhumanbodyforbettersegmentationPoseCut:Bray,Kohli,TorrModelsegmentationasBayesianlabelingproblemwith2labels:foreground,backgroundDetailsModelproblemasenergyminimizationproblem–modelasanMRFUseabasicstickmanmodelasahumanbodymodelAdaptivemodelforbackground
4、–GMMNeighbourhoodterms–GeneralisedPottsmodelMRF–MarkovRandomFieldsMarkovpropertyfortime:P(event:t)dependsoneventsattimesk5、model26degreesoffreedomGMM–GaussianMixtureModelModeleachpixelofimageasaweightedsumofGaussianfunctionsAdaptfunctionsusingeachnewframePixelmatchesexpectedvalue–background,elseforegroundExecutiondetailsForeachframeCalculateweightsforGMM,PottsmodelForgivenvalueof26vector(ba6、sedondegreesoffreedomofstickmanmodel)calculateenergycostforstickmanmodel(bydistancetransform)MinimizeenergyforBayesianlabelingbygraphcutMinimize26vectorbyrepeatedgraphcutsbyPowell'salgorithmSampleresultsA–originalframeB–segmentationbycolourlikelihoodandcontrasttermsC–wh7、enGMMtermsaretakenD–withposepriorcomponentsE–deducedposeComparisons
5、model26degreesoffreedomGMM–GaussianMixtureModelModeleachpixelofimageasaweightedsumofGaussianfunctionsAdaptfunctionsusingeachnewframePixelmatchesexpectedvalue–background,elseforegroundExecutiondetailsForeachframeCalculateweightsforGMM,PottsmodelForgivenvalueof26vector(ba
6、sedondegreesoffreedomofstickmanmodel)calculateenergycostforstickmanmodel(bydistancetransform)MinimizeenergyforBayesianlabelingbygraphcutMinimize26vectorbyrepeatedgraphcutsbyPowell'salgorithmSampleresultsA–originalframeB–segmentationbycolourlikelihoodandcontrasttermsC–wh
7、enGMMtermsaretakenD–withposepriorcomponentsE–deducedposeComparisons
此文档下载收益归作者所有