Generation of optimized facial animation parameters

The MPEG-4 FBA (Face and Body Animation) standard provides a set of FAPs (Facial Animation Parameters) for animating a talking face with its moods and expressions. A face model driven by this set of FAPs can produce high quality animation at a bitrate as low as 2 kb/s. The applications for the face model range from as diverse areas as video phone and video conferencing for wireless, portable units like PDAs (Personal Digital Assistants) and cell phones, to game development and movie production. The success of facial animation relies on being able to track facial features and to generate FAPs accurately and reliably.; This thesis presents a method for extracting FAPs from a person's face in a video sequence. The goal is to generate FAPs that make the animated face resemble the original face in the video sequence. The proposed method is based on feedback from the render unit during the FAP generation process. This ensures that the animated face is as close to the original face as possible. A penalty function is derived to measure the resemblance between the animated and the original face. The FAP generation process includes optimizing the penalty function, which consists of a match function and some barrier functions. The match function compares how well an animated face matches the original face in the video sequence. Knowledge about the appearance of a normal looking face is contained in the barrier functions. Each barrier functions indicates the level of distortion from a normal looking face for a certain part of a face, and advises the optimizer. Unnecessary FAPs are eliminated and the search space is partitioned into smaller, independent subspaces to speed up the optimization process. Steepest Descent Method, Cyclic Coordinate Method, Linear Search as well as Golden Section Line Search are applied to obtain an optimal solution.; The results show that the generated FAPs are accurate and the proposed method is very robust. The generated FAPs can drive animations that are lifelike and truthful to the original sequence, making them suitable for very high quality applications.