Mamba-based 3D Human Pose Estimation Algorithm for Track and Field Training

Abstract

In track and field training, the accuracy of athletes' technical movements directly affects their sports performance and sports injury prevention. Traditional posture analysis methods mainly rely on the coach's empirical observation, but this approach suffers from the shortcomings of limited evaluation angles and untimely feedback, which makes it difficult to provide accurate and real-time movement optimization especially in high-speed and high-dynamic sports scenarios. To address this problem, this paper proposes PoseFusionNet, a 3D human posture estimation model that combines OpenPose and Mamba algorithms. The model maps the 2D joint data extracted by OpenPose to 3D space and combines the timing information to improve the posture estimation's accuracy and robustness. Experimental results show that PoseFusionNet can still maintain high accuracy under multiple challenging scenarios (e.g., high-speed motion, occlusion, and illumination changes) and has real-time feedback capability, which can satisfy the dual demands of accuracy and real-time performance in track and field training. The model provides a new technical tool for intelligent sports training, which can help to improve the training efficiency of athletes and reduce sports injuries.