Tracking-by-selection pose estimation in videos
Jing Xiao · 2016
In this thesis, we propose a novel, effective, yet simple tracking-by-selection algorithm for human pose estimation in videos. The problem is solved by two steps, firstly, top N pose candidates are generated from each frame of the video; secondly, frame to frame temporal smoothness between poses across different frames are guaranteed by selecting the trajectory with the least Nuclear norm of its Hankel matrix among all the possible combinations. In the end, we also discuss the necessity of cleaning the selected trajectory by rank minimization to remove the effects of noise and outliers. Our dynamic based approach not only exhibits the ability to select the smooth trajectory from N-best detections accurately in the MoCap dataset, but also finds its value in video based tracking-by-selection human pose estimation framework.