Motion Capture Data based Animation
Although motion capture systems provide a reliable way to acquire realistic 3D motion data, motion capture data is hard to edit and adapt to new requirement. My research features an emphasis on designing tools for animators to produce animation with motion capture data at an approximately semantic and intuitive level. For this sake, wavelet analysis, Markov process and graph theory are introduced into motion synthesis to analysis a single motion or motion library.
Wavelet Based Motion Processing and Space-time Rectification: 
Inspired by multi-resolution analysis for image processing, we decompose motion capture data into multi-resolution levels with wavelet analysis. The coarse level represents the globe pattern of a motion signal while the fine levels describe the individual styles. Special motion style can be highlighted through enhancing the content at the corresponding level and can be fused into other motions by texturing them with related fine levels. And multiple motions can be synthesized by multi-resolution blending to create a series of new motions somehow like to the blended motions. However, original motion capture data implicitly preserve constraints to keep realistic, while the above manipulations may violate some constraints and result in some unrealistic artifact. Space-time rectification is proposed to reserve the essential constraints.
Graph based Motion Synthesis from Example:
The purpose of this work is to generate a motion for an intelligent avatar to move along a planned route using a pre-existing motion library. We consider each motion to be composed of a series of motion primitives, the minimal element that embodies the dynamic of a motion, and model it as a first-order Markov process. Then a directed graph, called motion graph, can be constructed from the motion library, with each motion primitive as a vertex. Within this motion graph, the desired realistic motion for avatars can be synthesized through a two-stage process: search a motion path within the motion graph through solving an optimization problem by transforming it into finding a shortest path in a hierarchy infinite directed graph using an adapted Dijkstra algorithm, and joint the motion path and adapt it to the route through a frame-by-frame strategy. Since each new motion is created by segmented motion capture data (motion primitives), the reality of each motion is preserved. And as motion primitives are used in synthesizing new motions instead of full-length motion clips, the promising motion space is well enlarged.
The drawback of this technique is that motion primitives are used to construct the motion graph directly, which results in the problem of high computational overhead in searching the optimal motion path when the motion library is large. To solve this problem, we propose motion cluster, a set consisting of motion primitives sharing similar dynamics. A new graph is constructed with each of its vertices as a motion cluster. Within this new motion graph, new application, random sampling for crowd simulation, is introduced. In this special application, a series of similar yet not same motions can be derived from a single motion cluster path through random sampling in each cluster. Because all the motion primitives in the same motion cluster share similar dynamics, these motions look similar to each other. As each motion is composed of a unique motion primitive sequence, it has its own characteristic. Each motion is formed using original motion segments (motion primitives), so the realism is well preserved.
Feng Liu and Ronghua Liang. Motion Path Synthesis for intelligent Avatar. The 4th International Working Conference on Intelligent Virtual Agents, Kloster Irsee, Germany, September, 2003, LNAI 2792, Springer-Verlag Berlin Heidelberg. pp. 141-149. pdf
 Feng Liu, Yueting Zhuang, Zhongxiang Luo and Yunhe Pan. A Hybrid Motion Data Manipulation: Wavelet Based Motion Processing and Spacetime Rectification. IEEE Pacific Rim Conference on Multimedia 2002, LNCS 2532, Springer-Verlag Berlin Heidelberg 2002. pp.743-750. pdf