Enhancing 3D Human Pose Reconstruction from Depth Sensors: Uncertainty-Constrained Bayesian Ridge Regression Approach

Human posture recognition plays a vital role in diverse healthcare applications. Depth sensor-based techniques provide an accessible and cost-effective solution for continuous motion capture, yet current pose reconstruction for dynamic movements remains challenging due to occlusion artifacts and inherent sensor noise. While existing methods focus on stationary or small-scale motions, they often lack generalization to real-world locomotion patterns. This paper presents an uncertainty constrained Bayesian Ridge Regression (BRR) approach to enhance depth-based motion capture accuracy. Our approach first establishes an initial reconstruction by mapping 3D skeletal data from depth camera to reference optical motion capture (MOCAP) trajectories through BRR modeling, explicitly quantifying prediction uncertainties. A subsequent dynamic optimization phase integrates anatomical constraints, temporal smoothness, and reliability weighting to refine predictions. Experiment results from a real-world dataset demonstrate a 42.12% reduction in average joint position error per frame compared to raw depth camera data, and consistently yields superior performance compared with all competitive methods. The approach also generalizes effectively to non-standardized dual-task scenarios, highlighting its robustness to gait variations. Overall, our developed approach shows significant potential in real-time high quality human posture recognition from depth camera.