Human Motion Segmentation And Annotation Based On Motion Capture Data

Since the1990s, with the development of motion capture technologies and the proliferation of motion capture devices, tremendous Mocap data has been generated. However, as the size of Mocap database increases rapidly and the type of motion becomes complex and diverse, it is more significant to manage and reuse the Mocap data in an efficient and rational way. The most acceptable way is classify different motion data by sematic. Therefore, segmentation and annotation of motion capture data has become a hot and challenging research field.This thesis focuses on human motion behavior segmentation and annotation technology. After reading series of related works, a LOF-based motion behavior segmentation method and a Motion Templates-based motion behavior annotation method are proposed. Besides, an automatic motion behavior segmentation system is designed and implemented. The main work of this thesis includes the following three aspects:In motion behavior segmentation, a novel segmentation method based on Local Outlier Factor is proposed. Local Outlier Factor, an anomaly detection algorithm, is introduced to find the transitions of two kinds of motion. Firstly, bone angle is employed to represent the motion characteristic in preprocessing stage. Secondly, sliding window is used as the statistical unit and statistic histogram is built for all bone angles in sliding window. Thirdly, LOF is calculated for each sliding window. Finally the segmentation point can be obtained by locating the peak of LOF curve. Comparing with the PPCA based segmentation method, the feasibility and superiority of this method has been validated by experimental results.In motion behavior annotation, a novel annotation method based on Motion Templates is proposed. Motion Field, a motion synthesis method, is introduced to synthesize motion template for each kinds of motion, because Motion Field synthesize each motion frame by compute the weighted average result of the k nearest neighbor of last motion frame. Thus, the synthesized Motion Templates, which capture the essence of an entire class of logically related motions, can be used as the judgment criteria for motion annotation. After obtaining the Motion Templates for all kinds of motion, the DTW similarity distance is computed between the unknown motion and all kinds of Motion Templates. Thus the unknown motion can be annotated by the smallest DTW result. An automatic motion behavior segmentation system based on OpenGL is developed. One of the main functions in this system is playing the Mocap data in GUI. The other is segmenting the Mocap data by LOF based method. The system, which is programmed by C++and is built with Qt Framework, is the foundation for future research work.