Spatiotemporal graph convolutional network for robust three-dimensional worker pose estimation from single-camera video with application to real-time ergonomic risk assessment

Accurate three-dimensional (3D) human pose estimation from single-camera two-dimensional (2D) videos remains a fundamental challenge in visual computing, particularly in industrial environments where occlusions, motion blur, and varying illumination are common. This paper presents a spatiotemporal graph convolutional network (SGCN) that addresses these challenges by jointly learning spatial and temporal dependencies from 2D keypoint sequences. SGCN integrates a Gaussian-based gated recurrent unit and a variance-based attention mechanism to robustly estimate 3D joint positions even when keypoint data are noisy or partially missing. The authors evaluate the method on standard pose estimation benchmark datasets (MPII, COCO, UP-3D, Human3.6 M) and real-world industrial worker dataset. SGCN achieves state-of-the-art mean per-joint position error (43.7 mm on fused data) while requiring only 4.2 Giga floating-points and 15.8 M parameters. As a visual computing application for industry, we demonstrate how the estimated 3D poses enable automated ergonomic risk assessment in real time. The robustness to occlusion and low computational cost make SGCN suitable for deployment in industrial settings, with potential extensions to biomedical gait analysis and artistic motion capture.