Research And Implementation Of Multi-view Video-based Human Motion Capture

The target of multi-view video-based human motion capture is to obtain sequences ofmoving human motion in multi-view video. This study is one of the hot topics in computervision research, which has important applications in human-computer interaction, sportsanalysis, animation, surveillance and so on.This thesis mostly focuses on the problems of markers automatic initialization, markerstracking, the three-dimensional of markers calculation and post-processing of markers, andthe main contributions are as follows:Firstly, to solve the problem of human body markers automatic initialization, this thesisemploys a method based on combination of predefined posture of performer and markersposition. Experimental results show that, the algorithm can initialize the human body markersautomatically.Secondly, tracking human body markers, this thesis proposes an approach to trackmarkers using two-dimensional image markers tracking and three-dimensional space markerstracking. In the method of tracking markers on the two-dimensional image, has two parts:regions tracking and human body markers tracking. Firstly, we initialize the model of thehuman body, and divide the body into three fixed regions; then, regions are tracked bytemplate matching method; finally, markers tracking can be completed combining with therelationship between regions and markers. The markers tracking process also utilizes theKalman filtering to predict the position of markers. In the method of markers tracking inthree-dimensional space, adding the epipolar constraints, video continuity constraints, humanbody joint length constraints and human joint angle constraints at the same time to guidemarkers tracking.Finally, smoothing and adjusting the markers’ three-dimensional data. In order to reducethe marker’ jitter which is caused by computational process errors and noise, and thediscrepancies between the actual length of human body joints and extracted length. This thesis adopts data smoothing and adjusting the position of markers to get more real andnatural motion sequences of performer.