A 3D Human Pose Estimation Algorithm Based on Feature Enhancement
Aiming at the problem that the estimation results of 3D human pose estimation networks based on hybrid methods are wrong in the face of occlusion and small targets,a 3D human pose estimation algorithm based on feature enhancement is proposed.In this network,multi-channel and multi-scale pyramid feature capture module is used to improve the feature extraction capability,and the graph attention module is used to refine the extracted features.Through the above two-step method,the feature refinement of the three branches can be improved,and the problem of inaccurate estimation in the attitude estimation process can be improved.After training on multiple datasets and testing on 3DPW datasets,the results show that compared with the existing advanced network ROMP,MPJPE and PA-MPJPE reduce the error of 3D human pose estimation process by 1.9%and 3.1%respectively.
万方数据
维普
