Limb-Constraint Based Spatial-Temporal Collaborative Network for Enhanced 3D Human Pose Estimation

Recently, methods that combine Graph Convolutional Networks (GCNs) and Transformers have demonstrated superior performance in 3D human pose estimation. However, they merely fuse captured local and global features simplistically, failing to reasonably constrain the inherent limb structure information of the human body. In this paper, we propose a novel spatial-temporal collaborative network architecture that acquires the positional information of 3D human pose through GCNs and Transformer pathways, and integrates a limb constraint module to correct errors caused by the human body's spatial-temporal variations, thereby enhancing the accuracy of 3D human pose estimation. We evaluate our model on two popular benchmark datasets: Human3.6M and MPI-INF-3DHP, achieving a significant improvement in the accuracy of 3D human pose estimation, which proves the effectiveness of the proposed method.