Motion synthesis based on dimensionality reduction

In this paper we propose a method for human motion synthesis based on dimensionality reduction. The human motion can be considered as lying on a low dimensional intrinsic structure because of the coordination between the body parts, such as the arms and legs. Based on this assumption, new motion sequences can be generated from motion capture data by learning linear dynamic models of motion segments on the resulted space of dimensionality reduction. The synthesized motions are compared with the original motion, and good performance is observed.