AN AI TABLE TENNIS STROKE MOTION ANALYSIS AND SCORING SYSTEM USING THE MULTI-HYPOTHESIS-GENERATION-BASED 3D HUMAN POSE ESTIMATION METHOD

This study employs advanced computer vision and deep learning techniques to analyze table tennis players' stroke motions, with a primary focus on 3D human pose estimation.  In this paper, we first proposed a deep learning model based on a multi-hypothesis generation transformer network for 3D human pose estimation. The model generates multiple possible 3D pose hypotheses from single 2D observations using 2D pose data detected from consecutive frames. Furthermore, we applied the proposed 3D pose estimation method to analyze the table tennis stroke movements, offering comprehensive evaluations and targeted training guidance. Two cameras were positioned to capture the athlete's stroke from both frontal and lateral perspectives, providing synchronized video data from both angles. Using our 3D human pose estimation method, keypoints were identified, enabling the calculation of five critical indicators: arm swing angle, knee bending angle, hip rotation angle, and stroke speed. We then calculate the similarity between player movements and a standard model using the dynamic time warping. On the public dataset Human3.6M, our method achieved an average joint error of 40.75mm, which is an improvement of 11.2mm over the classic pose former, and an improvement of 0.26mm compared to the contemporaneously outstanding D3DP. The contribution of this paper lies in improving the conversion of human pose from 2D to 3D and providing a more accurate 3D human pose estimation method. And the proposed system achieves quantitative analysis of batting techniques but also provides specific insights for technical improvement.