Fusion of Dual-layer Semantic Information for 3D Human Pose Estimation Network

When fusing multiple feasible solutions of human pose, existing methods do not adequately learn the dependencies between hypotheses, which easily leads to poor accuracy of the fusion results. Therefore, a 3D human pose estimation network fusing dual-layer semantic information is proposed. A hierarchical feature extraction module is proposed firstly to model the intrinsic structural information of human joint points, extract hypothetical features containing different levels of semantic information, and improve the utilization rate of position information of joint points. In order to further improve the network performance, a feature refinement module is designed secondly to transfer self-information of the hypothetical features, thus enhancing the correlation between joint positions. Finally, a hierarchical feature fusion module and an association calculation sub-module are proposed to learn the dependency relationship between multi-hypothesis features, and according to the relationship, cross-hypothesis information transfer is carried out between hypothesis features to fuse them into accurate and unified hypothesis features. Therefore, the different levels of semantic information of different hypotheses are fully utilized to obtain the final 3D human pose estimation results. The performance of the proposed model is verified on the Human3.6M, MPI-INF-3 DHP and HumanEva-I datasets respectively, and the experimental results show that the proposed method can improve the accuracy of 3D human pose estimation, and effectively deal with the cases of human self-occlusion and complex poses.