3D Human Pose Estimation and Temporal Kinematic Constraints for Dynamic Occlusion in Basketball

Abstract

The accurate estimation of three-dimensional human poses in  dynamic, multi-agent sports environments remains a significant  challenge in computer vision. Basketball, characterized by rapid  player movements, frequent inter-player contact, and severe dynamic occlusion, presents a uniquely difficult scenario for  standard pose estimation frameworks. This paper introduces a novel  methodology that integrates temporal kinematic constraints with deep learning-based pose estimation to mitigate the degradation of accuracy during occlusion events. By modeling the biomechanical limitations of human joint articulation and imposing temporal consistency across video frames, our approach reconstructs missing or noisy keypoint data with high fidelity. We employ a two-stage architecture that first lifts two-dimensional detections to a threedimensional space and subsequently refines these estimates using a temporal optimization modulegroundedinNewtonianmechanics.Experimental results on varying basketball datasets demonstrate that this method significantly reduces the Mean Per Joint Position Error (MPJPE) compared to state-of-the-art baseline methods,particularly in scenarios involving heavy defensive congestion. The integration of kinematic velocity and acceleration constraints ensures that the generated poses are not only visually plausible but physically valid.