[home]
Real-time Computer-synthesized realistic human motion

Goal: The Holy Grail of computer graphics - have a computer generate human animation that is indistinguishable from true human motion.

When Enrico Di Bernardo and I were developing models of human motion to make human detection and tracking more robust, we realized that our models could also be used to synthesize human motion.

Our method has the following properties:
- Motion models are learned from examples of real people moving
- Animation can be scripted with a high-level description (ex: walk along a certain path with a certain mood)
- Motion computation is very efficient, and can be done in real-time
- The resulting motion is quite realistic-looking (but we haven't found the Holy Grail just yet..)
Related Publications:
- Automatic Estimation and Synthesis of Human Motion, Luis Goncalves, Ph.D. Thesis, California Institute of Technology, March 2000
- US Patent # : (TODO: ask Enrico what the patent number and grant date are)
- "The Machine Stares Back": an online article in Caltech's E&S Magazine about this project and others in the Vision Lab

Sample Animations:
After learning the respective motion models, the animations were generated by specifying a sequence of directions in which to look or zap.

Look Motion Shazam-Zap Motion
Look around (2.2Mb)
Shazam-zap (2.6Mb)

Link to an old web page with another example here.
(a real-time DirectX demo is also available if you contact me.)

[home]