The Research On Multi-view Markless Motion Capture Based On Main Vector

3D human motion capture based on vision is a hot and difficult problem in thefield of computer vision, in recent years it also has been a forward direction in3Danimation, film and game. Multi-view3D human motion capture involves many keytechnology such as multi-view cameras calibration, image segmentation,3Dreconstruction, motion initialization and motion tracking and it becomes the researchfocus of domestic and foreign scholars.in the research of markless motion capture, model-based method has the problemof human model building, single motion type initialization and hard tracking;Learning-based method has the problem of needing to train sample library which iscomplicated to build. Due to the limit of the library capacity, we can not get thematched motion data entirely accurately. In order to address these problems, amodel-free method which can reduce the system complexity and improve the systemspeed would be proposed, the major work of this paper is as follows.Firstly, in order to address the problem of single motion type initialization andgetting the joint points by comparing voxel data to human model, an unconstrainedautomatic initialization method is proposed. Firstly, do spatial morphologicaloperation on the initial motion data, and then get the feature vectors of initial state, atlast combine these vectors to estimate space coordinates of each human joint position.This approach which recovers the3D posture by the information of vector realizeseasily, reduces constraint of posture and enhances the initialization speed.Secondly, because of the problem of human body high degree of freedom andocclusion and self-shielding phenomena, this paper proposes a method based on themain direction of motion tracking algorithm which combines the redundantinformation of reconstruction data, transforms the conventional motion capture’s jointtracking problem into a direction tracking problem. Firstly, this method uses temporaland spatial continuity of human motion, labeling the head and limbs voxel data bytemplate iteration and the shortest Euclidean distance criterion, and then get the maindirection vector by PCA algorithm, at last according to each direction vectorinformation, starting from the head position we can calculate all the remain jointpoints’ coordinates.Finally, In order to verify the above algorithm, this paper proposes own algorithm design on motion initialization and motion tracking in multi-view marklessmotion capture, and also has comparative error analysis with the real data which gotby algorithm of foreign scholars. The experimental results show that the presentedmethod in this paper can get high robustness, low complexity and also has a certaintheoretical and practical value.