Abstract：This paper presents a method for 3D human pose estimation using shape and motion information from multiple synchronized video streams. It separates the whole human body into head, torso and limbs. The state of each part in current frame is predicted by motion information, and the shape information is used as detector for the pose. The use of complementary cues in the system alleviates the twin problem of drift and convergence to local minima, and it also makes the system automatically initialize and recover from failures. Meantime, the use of multiple data also allows us to deal with the problems due to self-occlusion and kinematic singularity. The experimental results on sequences with different kinds of motion illustrate the effectiveness of the approach, and the performance is better than the Condensation algorithm and annealing particle filter.
2011, Vol. 33 
