Human Body Segmentation And Pose Estimation In Static Images

In this paper, we try to segment and estimate the pose of human body in static images with complex background. As there are large appearance variations caused by pose, clothing, illumination and view point (such as a standing man and a sitting man). All of these bring in great difficulties and challenges for the segmentation and pose estimation.We propose a new method for human body segmentation and pose estimation with adaboost classifier and Pictorial Structure (PS). We solve the segmentation and pose estimation problems in a way of mutual promotion which combines segmentation and pose estimation into one model.The human body can be modeled by the tree structure. The tree has ten nodes which represent each part of the body (head, torso, left upper arm, right upper arm, left lower arm, right lower arm, left upper leg, right upper leg, left lower leg, right lower leg). We train a template for each part using CRF model. The problem is solved by the Belief Propagation algorithm. Firstly, the message passes from the child nodes to the parent nodes until it reaches the root of the tree (torso). Secondly, the message passes back to all the child nodes from the root.We solve the problem in two main steps. Firstly, edge feature is used to build the template for each part and we can get an initial result of the probability of distribution of every part. Secondly, region feature is extracted from the initial segmentation image in the first step to build the region model for pose estimation and segmentation. In this way, we can combine the edge and region features together. Also we can combine the segmentation and pose estimation in one model.The dataset we use in the paper is well known in the field of pose estimation. It's a very challenging dataset. There are lots of human bodies with various poses in it. There are 305 images totally and 100 images are for training while the other 205 images are for testing. We get good segmentation and pose estimation results in this dataset.The edge and region features are very effective in the human body segmentation and pose estimation tasks. As we do not use skin color or other human-specific features, our method can be easily extended to other articulated objects'segmentation.