3D Human Pose Estimation Using Improved Semantic Graph Convolutional Based on Fusing Non-local Neural Network and Multi-Head Attention

Although semantic graph convolutions networks can effectively learn the dependencies between joints and bones, their accuracy in estimating human body coordinates is not high. Aiming at solving the above problem, this paper studies semantic graph convolutional networks and discovers the limitations of capturing complex long-range dependencies and assigning appropriate importance weights across graph nodes. To overcome these issues, a novel module, NMHA, is built by fusing multi-head attention and non-local neural networks to enhance the relational modeling capabilities of semantic graph convolutional networks. Furthermore, this paper proposes a new 3D human pose estimation model, NMHA-SemGCN, which incorporates NMHA to better address the defects of human pose estimation. Detailed experiments conducted on the Human3.6M and HumanEva-I datasets reveal that NMHA-SemGCN achieves significant improvements in accuracy over the previous approach. These results show the effectiveness and innovation of our method. Moreover, the paper presents a comprehensive approach for estimating human poses from monocular images to 3D skeletal coordinates utilizing the NMHA-SemGCN model, demonstrating its potential for practical applications.