Human pose estimation with heatmap-guided connection

Abstract

Human pose estimation was an important direction in the field of computer vision, which is widely used in human activity recognition, human-computer interaction, etc, but the accuracy of human pose estimation methods is usually poor. Based on this, a human pose estimation method based on heatmap-guided connection (HGC) was proposed to improve the accuracy of regression while maintaining low complexity. HGC method used the heatmap of keypoints to guide the regression of key points, and adopted a scale adaptive heatmap estimation to deal with the scale diversity of human instances. Then it constructed a scoring network combining the final posture structure and the heat value of keypoints to achieve good fitting effect between the estimated posture and the real posture in the dataset. The experimental results show that the average precision score of HGC is 72.9 on the COCO dataset, superior to the current mainstream algorithm.

Keywords

• human pose estimation
• heatmap-guidance connection
• keypoint detection
• scale adaptation