Research On Modeling And Reuse Of Human Motion Capture Data

With strong visual sense of reality and high fidelity, human motion capture (mo-cap) data has been increasingly applied in animation, game industry and medical rehabilitation. It is considered as one of the most promising techniques in animation producing and video games. Recently, a large amount of mo-cap data have been accumulated due to the growth of related researches and applications. If it is possible to effectively reuse the existing data to create new motions instead of repeatedly capturing ones, the cost of time, money and labor will be drastically reduced. On the other hand, by leveraging the existing mo-cap data, one can even create motions that are impracticable for a real man. Most importantly, it lowers the threshold of creating animations and games, which makes ordinaries can also do such jobs. Actually, mo-cap data reusing has already become one of the most active research topics in computer animation and graphics.In mo-cap data reusing, motion segmentation, retrieval and synthesis are the essential steps. In this thesis, our research mainly focuses on these three steps, and the major contributions are listed as follows:(1) A generic human motion representation-text-like motion representation is proposed, which is concise, easy to understand and has strong extensibility. This representation successfully builds a connection between mo-cap data and text data, making it possible for any text-based algorithm and model to be conveniently applied to mo-cap data, so that the human motion can be better processed and analyzed. Taking this representation as a foundation, we established unigram and bigram vector space models for human motions and proposed a moving range based word frequency adjusting method. Our model is able to accurately retrieve logically similar motions from the database.(2) Through textualization of mo-cap data and unsupervised topic modeling, we successful- ly discovered some intrinsic motion regularities:motion vocabulary and motion topics. Based on these motion regularities, a Local Semantic Coherence Curve is proposed to seg-ment long motion sequences into distinct behaviors. The proposed method not only takes full advantage of valuable information embedded in the existing motion data, but also has quite good extensibility of motion types. Compared with existing methods, our approach has better segmentation accuracy.(3) The motion parameters extracted by the existing parametric motion synthesis approaches are either structure-inconsistent or lack of intuitive meaning. In this thesis, we propose a sparse semantic parametric model, which is able to automatically extract meaningful mo-tion parameters that can directly control the jumping height, walking path and swinging range of arms, etc. The proposed model drastically reduces the operating complexity and causes less quality loss for synthesized motion. By adjusting the values of these parame-ters, desired motion can be easily generated in real-time.(4) In order to build stylistic human animation, a stylistic motion synthesis method is proposed based on Reconstructive ICA and Inverse Kinematics. The proposed approach avoids the manual predefining of the key values of independent components (or independent sub-spaces) and increases the quality of the extracted motion styles. By employing motion transition, the proposed approach is able to synthesize locomotions containing multiple styles and of arbitrary lengths. Moreover, the synthesis results can be refined by optimiza-tion with user specified pose constraints.