Motion Capture Data Based Human Motion Synthesis

Motion capture data based human motion synthesis is an important and difficult task in the field of virtual reality, which has been wildly used in nowadays CG movies and computer games. It focuses on reusing the existing motion data and improving the utilization of motion data for enriching the animation sequences. Considering the high expenses in motion capture hardware devices, motion capture data based human motion synthesis can greatly reduce the cost of animation creation and provide higher efficiency motion synthesis by computer automatic processing, which has emerged as a promising techniques for the automatic synthesis of realistic human motion.With studying the key technologies in motion capture data based human motion synthesis, a human motion analysis and synthesis system has been developed in this paper. Motion graph for motion synthesis are established based on the similarity computation within motion data sequences. The motion class templates based on motion feature analysis from motion capture data are constructed and implemented in the automatic motion class annotation of unknown motion sequences in database, which provide the semantic description for user interactive control on motion graph based human motion synthesis. Through the shortest path finding on the motion graph, combining the optimal motion transition generated, the new human motion sequences are created according to the user's requirements.First, the effective feature representations and extraction of human motion data are studied. In the low-level, the numerical motion features are extracted from the poses of different human motion states with the expanded data information within frames; in the high-level, the relational features are extracted to describe motion sequences which encode the certain geometric or directional aspects among the body's joints and provide the motion graph for synthesis with semantic information of motion features.Then, similarity computation of human motion data is studied. According to Euclidean distance, the similarity matrix is calculated between motion frames; the Dynamic Time-Warping algorithm is then implemented to find out the shortest path in the similarity matrix and get the similarity measurement between different length motion sequences.Next, automatic annotation of human motion sequences is studied. Based on the feature description extracted from motion data, the motion class templates are constructed through the self-learning process; the Dynamic Time-Warping algorithm is then implemented for calculating the local similarity matching between the motion templates and the unknown motion sequences for the automatic annotation process.Finally, motion graph based motion synthesis is studied and a human motion synthesis is developed. Motion graph is built up from the similarity analysis of motion sequences in database; Dijkstra algorithm is implemented to search the shortest path in the motion graph; and the motion transitions are generated by Geodesic distance measurement between frames,numerical analysis for deciding transition point and motion interpolation for smooth transition; Through annotations of motion sequences, the semantic description for user's interactive control to motion graph based synthesis is integrated; the system of motion synthesis can generate new human motion sequences which meet the requirement of users on the existing motion capture data.