Human Posture Analysis Under Partial Self-occlusion

Accurate human posture estimation from single or multiple images is essential in many applications. Two main causes of difficulty to solve the estimation problem are large number of degrees of freedom and self-occlusion. Treestructured graphical models with efficient inference algorithms have been used to solve the problem in a lower dimensional state space. However, such models are not accurate enough to formulate the problem because it assumes that the image of each body part can be independently observed. As a result, it is difficult to handle partial self-occlusion. This paper presents a more accurate graphical model which can implicitly model the possible self-occlusion between body parts. More important, an efficiently approximate inference algorithm is provided to estimate human posture in a low dimensional state space. It can deal with partial selfocclusion in posture estimation and human tracking, which has been shown by the experimental results on real data.