The Keck Laboratory for the Analysis of Visual Motion was established in 1997 through a grant from the Keck Foundation,. The Laboratory will be used to explore fundamental problems in the recovery of three dimensional models of human movements. The Laboratory will include a large set of digital color cameras that will simultaneously view a person perform an action. The video images from these cameras will subsequently be integrated into a dynamic three dimensional model of the activity. This model can then be viewed, from arbitrary perspectives, using advanced computer graphics techniques.
A suite of cameras collects synchronized videos of a patient undergoing rehabiliation after a serious accident. A therapist is able, subsequently, to view a three dimensional dynamic reconstruction of the patient's movements, and quantitatively compare her body movements to other patients at comparable stages of therapy. The same suite of cameras is used to watch a skilled technician conduct a detailed assembly or repair of a mechanism. Students remotely view the assembly process and personally control the viewpoint from which the process is observed. Developing exciting applications such as these required both fundamental advances in computer vision, as well as the construction of complex computer and sensing systems.
The research conducted in the Keck Laboratory addresses these problems in a fundamental way through the construction of dynamic graphical representations of human movement and human manipulation of physical objects. This will be accomplished by computer vision algorithms that
In Dr. Dariu Gavrila's Ph. D. dissertation we demonstrated how three dimensional models of a person being viewed could be employed to recover the time-varying articulation of the person in action.
These models were compared, using advanced optimization procedures, to between four and seven synchornized video data streams of the person in motion. The time-varying articulation of the 3D model was then superimposed onto the viewpoints to show the accuracy of the recovery process.
A paper describing this research appeared in CVPR '97 .
If you have comments or suggestions, email me at lsd@umiacs.umd.edu