2D Pose Estimation Based On Segmentation Consistency

Two-dimension human body pose estimation is to distinguish the position, orientation and scale of each body part in still image. First, each body part is detected in a image using part detectors, such as head, torso and so on. Then according to the kinematics relationship between every body part, all part candidates can be combined optimally and finally the best configuration is found as human pose. Human model can be mapped into a pictorial structure: each vertex represents a human part; each edge describes the relationship between a pair of parts. With Bayes framework, the evidence of each part can be considered as likelihood and the human kinematics is treated as prior, so human pose estimation problem can be changed to a maximum a posteriori (MAP) problem. Pose estimation which is the base of understanding human action belongs to image analysis. It is an important issue for intelligent processing and has great applications in artificial intelligence. By estimating pose, the scene of human behavior can be understood and timely responded with a machine.In fact, there are a lot of background noises, various human pose in image. Besides, the existing detection technique is limited. Those make pose estimation problem more difficult. For example, "double-counting" problem, that left and right legs overlap and "dismember" problem that the body parts are not compact enough. When two people are close in a image, it will confuse the ownership of their own body parts.This thesis presents an effective method that can realize the joint optimization of human segmentation and pose estimation, where human segmentation guides pose estimation and pose estimation can promote segmentation. First, we use Adaboost classifiers to detect each body part in each sampled positions and orientations in a image, thus obtaining the score of each part. Secondly, according to the human kinematics, we compute spatial and angle distribution of each part. Then, we perform message pass along the tree structure using the local features and the priori above via sum-product algorithm. Therefore, the posterior probability map is calculated. Next, we sample N largest candidates of each body part in the posterior probability and compute each weight that describe how it is to be the part by the region and probability of boundary. For bottom up cues, we over-segment the image to get a series of super-pixels and calculate the similarity between each pair. With the candidates we select, it is easy to estimate the distribution of foreground. Finally, with each body part weight, their connection relationship, the super-pixel foreground probability and the similarity of each super-pixel, we can build a linear energy function, and solve the optimization via integer programming. So, the human pose and segmentation are both achieved.The main contributions of this thesis are summarized as follows:(1) Based on the traditional sliding window detector, we incorporate the information of the connection relationship of each body part and detect each body part with a PS-based body part detectors, improving the detected results; (2) we propose a segmentation and pose estimation simultaneously optimal method; (3) We solve the double-counting and dismember problem, and test our method in the public dataset (Parsing, containing 305 human full body, in which 100 are used as training and 205 for test; Buffy, containing 268 human upper body). The results show that proposed method is very efficient.