Vector Hourglass Network for Human Pose Estimation based on Deep Learning
Huimin Li · 2019 IEEE 2nd International Conference on Automation, Electronics and Electrical Engineering (AUTEEE) · 2019
Human pose estimation (HPE) is the basis of human action recognition and analysis. It is better to locate the key points of the human body more accurately to achieve the desired results in certain applications. The paper will discuss and explore it. The overall pipeline refers to the method AlphaPose. Among them, the single person pose estimation network uses the proposed improved hourglass network - the proposed method calls it Vector Hourglass network. The main process of Vector Hourglass network is as follows. First, the heatmap values of each graph obtained from the first-level and the second-level hourglass network are saved respectively. For the heatmap values obtained from the first-level network, the geometric center of gravity of each obtained joint point's heatmap is taken as the joint coordinates for calculating the PAF weight, and then calculate the PAF values. By weighting the heatmap value getted by the first-level network to the thermal value obtained by the second-level network, the new thermal value of the second-level hourglass network is obtained. In the test phase, since the method of this project can only detect the pose of a single person, LSP (Leeds Sports Pose Dataset) single person sports dataset is selected as the test set from the existing open dataset. To verify the validity of the method, the project calculates the mean square error (MSE) and the mean absolute error (MAE) values of the heatmap that weighted PAF weights before and after. The results show that the positioning accuracy of key joint points is improved by 1.43 pixels in the image of 64×80 pixel size. The method applies the idea of intermediate supervision and part affinity fields (PAF) between the two levels of hourglass network to improve the positioning accuracy of the joints and makes the obtained human posture contour more obvious.