Human Pose Estimation On Template Matching

The purpose of human pose estimation in an image is to detect the position, orientation and size of each body parts, such as body torso, head and limbs. In non-finite scene, the missing of getting the information of limbs from an image often happens, due to the violation of background, the variation of human pose, the self-occlusion, etc. As a result, human pose estimation from a still image is a challenging work. This task is a crucial pre-requisite step to many high level vision applications, such as human-computer interaction and activity recognition. Therefore, pose estimation is a very meaning research in computer vision area.The mainly discusses in this paper is the use of template-based matching method to estimate the human pose in a still image. For reducing the cost of deformation when the neighbour-part Backtracking, this paper purposes a algorithm that using the head-torso poselet replaces the head-torso in the DPM. The method enhances the human model and renews the score function of pose, according to the width torso detected by the poselet and the appearance of part symmetry to constrain the parts. So the method can improve the detection of torso and limbs by enhancing the relation of parts and get the best result which is suit for the human model. It is important to notice that the method still based on PS structure with the tree-structure model can be efficient to compute in the phase of inference.In order to verify our algorithm has better applicability, this paper tests the method on the publicly dataset, such as LSP and PARSE. The result shows that the model has a strong ability to restrain the torso and limbs, thus improving the whole detection of pose estimation.