Research On Regression Algorithm Of Human Pose Estimation Based On Keypoint Matching

As a research hotspot in recent years,human pose estimation can be applied to every field of life.There are many problems in different application scenarios,such as the diversity of human pose,the occlusion of keypoint,the difference of target scale,and the background blur.Therefore,it is still a challenging task to achieve accurate human pose estimation.Existing deep learn-based multi-person human pose estimation methods are mainly divided into top-down and bottom-up methods,but most of the methods do not pay attention to the full use of local features in the network and how to accurately locate human keypoint.Aiming at the shortcomings of existing methods,thesis takes multi-person human pose estimation as the research starting point,and the optimization network model and keypoint heatmap regression algorithm as the main research clues to carry out the research work.The main research contents are as follows:(1)The human pose estimation framework based on hourglass network can realize the pose estimation of large target figures.However,in the real scene,there are some phenomena of unclear identification of small target people and occlusion of keypoint,and the model does not integrate local features in the context network when generating keypoint heatmap,which will lead to some small keypoints(eyes,ears,etc.)cannot be recognized.In response to these problems,thesis proposes a Attention Stacked Hourglass Network(ASHN)framework.ASHN firstly uses the keypoint location algorithm to detect and locate the keypoint of human pose.Using hourglass network stacking as the model for the second step of training.In thesis,attention mechanism is integrated into the hourglass network to consolidate local features and achieve accurate identification of keypoint.Then,Focal L2 Loss function was added in ASHN training stage to stabilize intra-class and inter-class unbalance of samples and improve the applicability of the model.The experimental results show that compared with SHN and G-RMI algorithms,the proposed model framework can effectively improve the efficiency of human pose estimation and recognition.(2)In the coordinate regression of keypoint of human pose,the heatmap of keypoint is usually generated by Gaussian function to obtain the corresponding keypoint.However,Because the standard deviation of different sizes in the Gaussian function will cause errors to generate the keypoint region,and can not be matched properly when the keypoint are grouped.Aiming at these deficiencies,a Adaptive Standard Deviation Heatmap Based on High Resolution Network(ASD-HRNet)human pose estimation scheme is proposed in thesis.In this scheme,HRNet and channel attention module are used to generate heatmap of keypoint.Then the ASD algorithm adaptively sets the size of standard deviation to adjust the heatmap of keypoint.Finally,the keypoint are grouped together.Experimental results show that compared with HRNet-W32 and HRNet-48 algorithms.When the scale difference of the target keypoint cannot be matched normally and the accuracy of human pose estimation is low,the proposed human pose estimation framework has higher recognition accuracy and stronger robustness.To sum up,aiming at the key challenges in the research of human pose estimation,thesis basically achieves the target human pose estimation with stronger robustness,higher recognition accuracy and lower complexity.