Inferring Human Upper Body Motion

We present a new algorithm for automatic inference of human upper body motion in a natural scene. The initial motion cues are first detected from the captured video. A graph model is proposed for human upper body motion, and motion inference is posed as a mapping problem between state nodes in the graph model and the motion cues in images. Belief Propagation and dynamic programming algorithms are utilized for Bayesian inference of upper body motion in this graph model, which captures constraint of human body configuration under specific view angles. A multi-frame inference algorithm is proposed to combine temporal smoothness constraints in human upper body motion. The algorithm is applied in a prototype system that can automatically detect and track human upper body motion from captured videos, without manual initialization of human body parts. We present evaluation results of our algorithm on this system and some further considerations for refine the model in the future.