Research On Huaman Pose Estimation In 2D Image Based On Pictorial Structure Model

As an important branch of computer vision,human pose estimation' is widely used in many fields,such as motion recognition,human-computer interaction,video analysis,behavior tracking and so on.Human pose estimation also faces many challenges,such as the high flexibility of human pose,the influence of complex environments and so on.The initial research of human pose estimation is to construct the human pose template library,and realize the mapping of image feature to the post in the database through the specific mapping relationship.In fact,the human pose in the image is ever-changing,and this method can only describe the partial pose of the human body,so the limitation of this method is large.In order to more accurately and accurately simulate the human pose map structure model,the model makes the human body as being composed of different parts,and the parts are connected to each other in a certain way,so the human pose estimation is changed from the overall detection to the small part detection.The advantage of this method is that it can simulate the human body in any posture without constructing a human posture library.At present,the research on human pose estimation mainly focuses on the pictorial structured model,in which the human tree model is a popular method,and it has the advantages of simplicity and efficiency.In this paper,we focus on the problems of human pose estimation in single image.And,based on the mixtures of parts model of pose estimation,a enhanced context aware model is proposed to improve the robustness of the human body model.The mixtures of parts model introduces the orientation information of parts based on the tree model,which can better capture the spatial relationship between human body parts.However,the mixtures of parts model,like the tree model,uses only the spatial constraints of its child parts during the detection,and ignores the constraints of its parent parts.Therefore,the model is prone to error detection under the influence of noises and large deformation of human body.According to the characteristics of information transfer from bottom to top in the tree model,this paper realizes the constraint of the upper part to the lower part by generating the virtual part of the human body,so as to reduce or eliminate the influence of noises,and improve the accuracy of the detection.Since the detection of human body starts from the lowest part of the leaf parts,its accuracy will affect the detection accuracy of the whole human body,In the tree model,the detection of leaf parts does not use any spatial constraint information.Therefore,we use a wider range of constraints to improve the detection accuracy of human leaf parts.Moreover,the model of this paper also maintains the basic tree structure,so it can be solved efficiently with dynamic programming.In addition,due to the influence of occlusion in human pose estimation,a new occlusion aware model based on the above method is proposed to solve this problem.Previous studies on occlusion mainly focused on other-occlusion,ignoring the self-occlusion of the human body.Our method divides occlusion into two cases:self-occlusion and other-occlusion.The self-occlusion of the human body is usually caused by the mutual occlusion between the symmetric parts,so the model adds constraints between the symmetrical parts of the human body and learns their occlusion relationship from the datasets.Finally,the effectiveness of the context-aware model and the occlusion-aware model is proved by experimental comparison on the popular datasets.