Research And Implementation Of Human Pose Keypoints Detection Technology

Human pose keypoint detection technology,also known as human pose estimation,solves the problem of locating the keypoint coordinates of person in 2D images or videos.As the basis of understanding human behavior in images and videos,human pose estimation has a wide range of applications in advanced human-computer interaction,intelligent security and other fields.This thesis focuses on the keypoint occlusion problem and the low efficiency problem in current human pose estimation.The main work is as follows:(1)Aiming at the problem of negative transfer effect of occluded points in the process of pose refinement,This thesis proposes to use the visible probability of keypoint which is predicted by occlusion information network to reweight the keypoint heatmap.Then send it to the pose refine network together with the original image,reduce the negative transfer effect of occluded points to ouput more accurate pose.Experiments were conducted on MPII dataset,The results show that the proposed pose refine network run as a type of post-processing method can be cascaded with pose estimation network to improve their prediction accuracy,and the proposed pose refine network can significantly improve the prediction accuracy of occluded points,the prediction accuracy of some keypoints is improved by more than 10%.(2)Aiming at the low efficiency problem of current pose estimation network,a new lightweight module CpConv is proposed.Based on the idea that the feature maps of complex network are redundant,the input feature maps are compressed by group convolution to obtain meta-features,and then the information between groups is fused on the basis of the metafeature maps,which achieve the purpose of compressing the network while ensuring the effective communication between groups.This module is embedded in the original pose estimation network as a plug-and-play manner to simplify the network.Experiments were conducted on MPII dataset,The results show that the improved network’s prediction accuracy only lost 2.5%,but the amount of parameters and calculations are reduced by nearly five times.