Human segmentation based on Bayesian model

In this paper, a novel method is proposed for human segmentation in the presence of occlusion or shadows. We utilize human body shape model to interpret the foreground in a Bayesian framework. In order to reduce process time, we use head-torso model instead of full-body human shape models in the solution space of Bayesian model. We obtain MAP (Maximum a posteriori) by using MCMC (Markov chain Monte Carlo). The choice of proposal density also can avoid miss-detections caused by mistaking small targets as image noises. In our method, we firstly detect head candidates; then, we construct Bayesian model and calculate MAP of it; finally, we select optimum human models for the candidates and segment human objects precisely. In a number of human segmentation experiments, our method not only achieves good performance in occlusion situations, but also is resistant to background noises.