A Semantic Web-Based Approach for Bat Trajectory Reconstruction With Human Keypoint Information

Abstract

Restoring the trajectory of a bat from a table tennis match video is critical in analyzing a table tennis technique and conducting statistical analysis. However, directly bat location detection in each frame is challenging due to changing shapes caused by varying movement directions and speeds, leading to ambiguity. This paper develops a novel two-stage method. The first stage utilizes YOLO for bat detection in each frame, followed by filtering out erroneous candidate boxes. In the second stage, the authors use a temporal prediction model that integrating human keypoint information and interpolation to reconstruct a complete bat trajectory with minimal errors. The method's effectiveness and performance are evaluated on our video datasets. The evaluation results demonstrate that the proposed method outperforms traditional methods on precision performance metrics. The error screening algorithm improves precision score to nearly 1. In addition, the method has the recall score 22.3% higher than YOLO 's and also 1.4% higher than that of YOLO with cubic spline interpolation.