Development of Multi-person Pose Estimation Method Based on PAFs.

Human keypoints are effective human pose descriptions. Human behavior can be recognized by the motion of keypoints of human bodies. In this paper, we propose a method, which is based on a PAFs approach, for human keypoints detection. The proposed method makes improvements in two aspects: (1) It perfects the joint points matching algorithm by re-matching. (2) It uses multi-branch PAFs to correct the keypoint connections, thus improving the wrong connection problem of upper and lower limbs for multi-person keypoint detection. The improved PAFs method, whose mAP reaches 53.6% on HKD dataset, improved the score of the original method by 2%.