EXTRACTION OF MPEG-4 FAP PARAMETERS FROM 3D FACE DATA SEQUENC ES

The communicative power of face makes modeling and tracking of the face an active research topic. The new MPEG4 standard includes support not only for natural video and audio, but also for synthetic graphics and sound. Representation and animation control of human faces are defined by deforming a generic facial model. We hereby use this standard as a testbed for our nonrigid motion analysis on the face data. Our research is focused on analyzing 3D facial sequences using invariant properties of differential geometry. Our main reasons for using 3D data input are as follows: i) they contain more geometric hints; ii) as in MPEG-4, facial parameter set is typically 3D based; iii) the extracted 3D facial motion can be used in other fields such as natural language understanding and medicine, where 3D analysis becomes crucial. Other related application areas include low-bandwidth video conferencing via virtual human, model-based compression, human-computer interface, and entertainment industry. Given 3D facial images, we utilize three methods integrated in a motion tracking algorithm that use i) discriminant changes, ii) unit normal changes, and iii) homothetic motion assumption. Our investigations include experiments on face at different areas of interest with various shapes and sizes of search windows to achieve better performance. Motion tracking results are converted into FAPs, as specified in MPEG-4 standard. Our preliminary experimental results look promising and they are demonstrated in tracked images and tables. Our future work includes automatic registration of FAP locations, adaptive selection of search window sizes and shapes and generalizing the function representing facial motion.