Human Pose Analysis Based On Cross-stage Deep Network

In the field of computer vision,estimating human poses from still images is an important and challenging task.The purpose of human pose estimation is to determine the pixel locations of body keypoints in a given image.Human pose estimation is the basis of many advanced computer vision tasks,including activity recognition,image retrieval,etc.However,due to the influence of joint complexity and mixed background,it is still a challenge to locate the human keypoints accurately.Based on the existing research on human pose estimation,a cross-stage structure is introduced to locate the body keypoints in the image,which improves the accuracy of keypoints locating.Moreover,we design belief inference based on the whole credible keypoints to improve the detect accuracy of human keypoints with poor results as a post-processing method.The main work of this thesis is as follows:(1)This thesis elaborates the basic process of human pose estimation in detail,introduces several classical human pose estimation models according to traditional models and deep network,and their respective advantages and disadvantages are analyzed,which lays a theoretical foundation for the development of the thesis.(2)A human pose estimation model with cross-stage structure is designed to improve the accuracy of detection.The model extracts the initial features through the first stage network and uses a multi-stage network to reflect the effects of features on keypoint locations at different scales.At the same time,a joint loss of multi-scale locating is designed to connect the initial features with each scale feature in a cross-stage way,which shortens the calculation path between the locating error at the output and the locating error at each scale.And it improves the effective learning ability of the initial features of human body keypoints.(3)Aiming at the problem of human pose estimation,this thesis designs a process of belief inference based on the whole credible keypoints for the reason of the difficulty in dealing with the poses in occlusion.According to the keypoint belief maps outputted by the model of human pose estimation,we can judge its visibility.It can be divided into two groups,one is credible and the other is incredible keypoint.Then the group of credible keypoint belief maps is obtained.According to the human Gaussian conditional distribution between the two different keypoints,belief inference is performed on the incredible keypoints caused by occlusion.The purpose of the inference is to achieve aneffective results on occluded poses and to improve the robustness of the model in occlusion.