Enhanced multi-branch network for 3D human pose estimation from single-frame skeletons

Three-dimensional human pose estimation is pivotal in diverse fields such as motion analysis, healthcare, and smart home design. This study addresses the challenge of deriving accurate 3D human poses from single-frame two-dimensional skeleton data. We begin by analyzing the error distribution of a foundational network for estimating human pose keypoints. Based on this analysis, we propose a torso-centered grouping method and design a multi-branch network for enhanced 3D pose estimation. Furthermore, we introduce a distance constraint loss function to mitigate the structural information loss caused by grouping and to reduce estimation errors at limb extremities. Our experimental results on benchmark datasets demonstrate that the proposed network significantly improves the accuracy of 3D human pose estimation. These findings hold promise for advancing applications in motion analysis and abnormal behavior monitoring.