Variable Length Hidden Markov Model for Motion Generation

Description

In this project, a motion learning and synthesis system will be developed for which the variable length hidden Markov model (VLHMM) is proposed. Compared with the well-known first-order hidden Markov model (HMM), the VLHMM is able to learn variable length of motions with accurate temporal prediction, which makes it possible for the VLHMM to discover basic elements of motion as atomic behaviours. As a hidden model, the VLHMM is more expressive than the variable length Markov model. Such a representation is compact for the high-dimensional degree of freedom (DoF) of the human body. With the ability to learn from captured 3D human motions, this system can generate a variety of realistic, natural-looking new motions. The proposed system will have a substantial potential in motion editing and synthesis. It will also lay a foundation for major initiatives to develop Motion Engine and Autonomous Animation for applications in computer animation, computer games, special effects, human-machine interface and surveillance.