Research On Touch-free Human Body Motion Capture Under Multiple Viewpoints

The goal of motion capture is to detect and record the motion and expression of moving objects, which can be represented as poses of the objects at any time, and converted to abstract digital format. Motion capture is useful to new-generation human-machine natural interaction, cartoon production, game production, motion analysis, virtual reality, etc. Because touch-free human body motion capture has many advantages, such as non-compelling, low cost, high automaticity, it is an increasingly hot research project in motion capture field. Still an open question is how to acquire pose information of human body from image sequences independent of special equipment and markers in the presence of noise, cluttered background, occlusion and self-occlusion. In the paper, a few relevant issues in human body motion capture under multiple viewpoints are researched, coming up with some new solutions and algorithms. Main content of this work is as follows:1. For extracting motion objects, an adaptive background model with hierarchical Gaussian Mixture Model (HGMM) is proposed. It enables motion capture in more cluttered scenes by solving three problems in adaptive background model: The background model was initialized too slowly; those objects that stop moving temporarily were absorbed into the background model too quickly; sharp illumination changes could not be handled. In this section, a learning algorithm for learning background model on small sample sets is introduced first, and then a new algorithm for selecting background model is designed. Finally, hierarchical configuration of background model is described.2. For extracting motion objects, a new Markov Random Field - Maximum A Posteriori method (MRF-MAP) with adaptive GMM background model is proposed, i.e. GMM-MRF. To extract moving objects accurately in real time, it solved four problems in existing MRF-MAP method: False positives caused by dynamic background could not be handled; Shadows were not well eliminated; extracted contours of moving objects did not match the real ones in the images; the optimization algorithm was inefficient. In this section, contributions are made to the MRF-MAP method at four aspects. Firstly, energy term based on adaptive GMM background model is designed. Secondly, shadow elimination term is designed. Thirdly, new...