The Research On Human Pose Recovery And Tracking From Multi-view Videos

Recent years, video based motion capture technology has been one of the hottesttopic in the area of computer vision. It is mainly used in animation and film,interactive games, motion analysis (sport, medical) and surveillance. The keytechnology of motion capture is the pose estimation, recovering3D human body poseparameters from2D videos. The main components of the study are camera calibration,image segmentation,3D reconstruction, posture initialization, motion tracking etc.The tracking is the key problem of the research.Most markerless motion capture methods can be divided into two categories:model-free and model-based. The model-free method need to build the trainingdatabase, so it has high complexity. But the model-based method has the problem ofmismatching from the model to the observation data. According to this issue, thispaper proposed a hierarchical method for human pose tracking, which can accuratelyrecover the body posture. Our main work is summarized as follows:First of all, this paper presented a method for automatic initialization. This methodcan automatically obtain the initial human3D posture from the initial frame throughintegrated the3D information and the2D silhouette information, and built an adaptivehuman model, thereby complete the model initialization. This method first extractsforeground information from the selected view, named as human silhouette, and thenthe region segment method is used to partition the silhouette. Then, the labelledvoxels can be obtained by using Shape-From-Silhouette (SFS) algorithm, combinedwith the body information of the2D image and the3D space information. Finally, theparameters of the3D human skeleton and its space information can be extractedaccording to the labelled voxels as well as the topology constraint of the humanskeleton. So, the initialization of the human model can be accomplished.Secondly, according to the continuity of human action in time and space, this paperproposed a method of motion tracking based on labelled voxel. In order to make themotion tracking more fast, we divided the voxel data into three categories: surfacevoxel, intrinic voxel, and the middle voxel. The middle voxels are used for motiontracking, and the surface voxels are used for global optimization. Firstly, this methodget the head information through the method of head template matching; Then, themain vector of the torso can be gained through vector predict. The root informationcould be extracted by using the Principal Component Analysis (PCA)method; After then, the remaining voxels are labelled to different parts, according to the constraintsof the Mahalanobis distance, and then matching the voxels of each part to the previousmodel part using the Iterative Closest Point (ICP) algorithm to gain the local optimumparameters. At last, we use the global optimization method to obtain the body postureof the current frame.Finally, this paper complete the algorithm of the motion initialization and motiontracking in this markless motion capture system, and realize the code validation ofthese algorithm. We also use the data and the error evaluation method of the BrownUniversity. The experimental results show that, this method is feasible, lowcomplexity, and has certain research value and application value.