Research on Human Pose Estimation Algorithm Based on Two-Stream Residual Steps Network

Abstract

Human pose estimation plays an important role in machine vision. 2D-based human pose estimation has received more attention due to its low 2D space complexity. However, many algorithm networks detect human key points in different environments, and their key-points prediction accuracy will decline to varying degrees. In order to solve the problem of weak generalization of human keypoints detection in different environments. First of all, this paper, based on the small influence of illumination in the MSR space image and the rich human detail information stored in the RGB space image, proposes a pose refinement network and a pose correction network to extract features from the two spatial images. Moreover, in order to better fuse the feature information extracted from the two complementary spaces of MSR and RGB, an adaptive feature fusion algorithm is proposed. Finally, keypoints are regressed by generating a heatmap. The experimental results show that the network has high prediction accuracy on both MS COCO and MPII datasets, and can maintain good generalization in different external environments.