A human pose estimation algorithm based on E-DCPose is proposed to address issues such as occlusion and motion blur in video human pose estimation.On the basis of the DCPose framework,pose semantic propagation and temporal heatmap alignment methods are introduced to alleviate the problem of inaccurate joint positioning caused by occlusion and motion blur during human motion using contextual temporal information.At the same time,by increasing the number of video frames to provide sufficient temporal information for the model,the accuracy of human key point detection is improved.This article conducted ablation research on the effectiveness of various improved modules of E-DCPose on the datasets PoseTrack2017 and PoseTrack2018,and compared experimental analysis with existing models.The experimental results show that the detection accuracy of E-DCPose is superior to all comparison models and significantly superior to the baseline model DCPose.
human pose estimationE-DCPosepose semantic propagationtemporal heatmap alignment
万方数据
维普
