Bayesian Estimation of 3-D Human Motion
|MERL Report: ||TR98-06: Michael E. Leventon, William T. Freeman
We address the problem of reconstructing the 3-dimensional motions of a human figure from a monocular image sequence. We take a statistical approach, and use a set of motion capture examples to build a gaussian probability model for short human motion sequences. We first study this model in a simplified rendering domain. This yields analytic results for the optimal 3-d estimate given a 2-d temporal sequence, as well as for which motion modes are difficult to estimate. The results from the simplified rendering conditions show that if we can overlay a stick figure on an image of a moving human, we can estimate his or her 3-d motion well. We built an interactive tracking system to process real video sequences, and can achieve good 3-d reconstructions of the human figure motion.