Vision-Based Realtime Animation Synthesis

This paper introduces the design of our real time vision-based motion performance animation system. The system requires user to wear a small set of markers. The low-dimensional control signals from user´s performance are first used to construct a series of local models, which could be later used to animation synthesis. When constructing local models, we preprocess motion capture data to K-nearest neighborhood graph and store these data in KD-tree to ensure model building is real-time. In animation synthesis, we use an approach named locally weighted linear regression to synthesis the animation data closest to current pose. The performance of the system is tested by having users do three kinds of motion: running, walking and jumping.