Thomas S. Huang
3D Motion Analysis and Model-Based Image Compression in High-Definition Television
The development of High-Definition Television (HDTV) technology will have a tremendous impact on the nation's economy in the next 20 years. Professor Huang plans to initiate a research project on image compression in HDTV. It is hoped that results from this project will serve as a basis for attracting future funding from government and industry to expand UIUC's involvement in HDTV research. Past experience in image processing and computer vision put UIUC in an excellent position to become a leading contributor to this important field.
Because of the high data rate of digitized HDTV signals, compression is often necessary. In many applications, very large compression factors are needed--possibly 5000:1 or more. Professor Huang and his team propose to take a model-based approach to achieve very high compression. They concentrate on the teleconferencing scenario where the scenes contain mainly face/head views. 3D static models of the face/head are stored at the receiving end. The transmitting end extracts the motion information of the speaker's face/head (from image sequences taken by TV camera) and transmits only this information to the receiving end. Then the receiving end uses this motion information to drive the 3D static model and, by applying computer graphics techniques, generates and displays image sequences representing perspective views of the moving face/head.
The most challenging aspect of this system is the motion analysis at the transmitting end. Thus, the major part of the team's research will be on the use of computer vision techniques to extract 3D face/head motion from image sequences. Three levels of motion descriptions will be investigated: Vertex displacements, action units, and events. A vertex is a key 3D point on the face. An example of an action unit is "tilt of head." An example of an event is "smiling." It is hoped that by using motion information at the event level, extremely high compression (100,000:1 or more) could be achieved.