Capture and synthesis of human motion in video sequences

We present a knowledge-based framework to capture and represent human walkers in video. The system models the human body as an articulated object of twelve rigid body-parts whose motions are almost periodic and subject to dynamic constraints. The resulting representation is compact and composed of the motion, shape, and texture for each of the body-parts. We apply the representation to regenerate the original sequence and to synthesize articulated 3D human actions.