Behavioral Segmentation And Motion Analysis For Human Motion Capture Data

Due to the popularity of optical motion capture system and the development of mo-tion capture technology, human motion capture data is growing huge and has been suc-cessfully implemented in many areas such as biological science, computer vision, com-puter animation and virtual reality. How to exploit the data reasonably and effectively has become more and more important. The current motion capture technology is an im-portant and hot task for the analysis, organization and administration of human motion capture data.This thesis focuses on studying behavioral segmentation and motion analysis in-cluding motion template extraction and motion annotation for human motion capture data, which provides the corresponding method for each problem, and evaluates every method by experiments. Research works of this thesis mainly include the following three aspects:For the behavioral segmentation of human motion capture data, a behavioral seg-mentation method based on Kernel Dynamic Texture (KDT) is presented. First, the ref-erence sequence clips and subsequence clips for similarity comparison are selected from the raw human motion capture data. Then, the selected reference sequence clips and subsequence clips are modeled by KDT based on kernel Principal Component Analysis (PCA) method. Finally, Martin distance is exploited to measure the similarity, and the behavioral cuts are detected based on the variation characteristics of similarity distance. Experiments show the good behavioral segmentation performance of our approach.For the extraction of motion templates, a new hierarchical Latent Motion Template (LMT) learning method is proposed. First, the motion templates of input human motion capture dataset are modeled by Hidden Markov Model (HMM). Then hierarchical LMT is learned with two-stage Expectation Maximum (EM) iterative algorithm. The biped locomotion LMT in the top level is extracted from the first stage. The motion class LMT in the second level and the motion style LMT in the third level are extracted from the second stage. Experiments show that our proposed method is robust for different styles and classes of human motion capture data.For the annotation of human motion capture data, a motion annotation method based on hierarchical LMT is proposed. For the annotation analysis of each level, the match-ing similarity between hierarchical LMT and subsequence clip is computed by Dynamic Time Warping (DTW) method, which includes the annotation with biped locomotion LMT in the top level, the annotation with motion class LMT in the second level and the annotation with motion class LMT in the third level. Experiments show that our pro-posed method can effectively accomplish the hierarchical annotation for the human mo-tion capture data.