부산대학교 도서관

Many studies in recent years have discussed the automatic tracking and analysis of ball sports. In table tennis (TT) competition, it is important to know who uses the artificial intelligence (AI) technology to accurately analyze the complex competition process in TT to manage the characteristics of small-size ball, speed and several rotation changes, leading to a winning opportunity. The purpose of this study was to determine the trajectory and bounce point of the ball movement through computer vision technology. We proposed a two-stage deep learning network algorithm for the ball tracking and bounce estimation from TT videos. The network's first stage detected the TT ball position in the image. In the second stage, we used obtained region of interest (ROI) image sequences from the first stage to predict the precise coordinates of the TT ball. Results showed that the precision and recall of ball tracking reached 95.51% and 92.65%, and the bounce detection reached 90.90% and 88.23%, respectively, when the film with a frame rate of 120 FPS (frames per second) was used. In the film with a frame rate of 60 FPS, the ball tracking precision and recall reached 94.69% and 76.80%, and the bounce detection reached 86.36% and 92.59%, respectively. These findings demonstrated the effectiveness of the proposed deep learning methods in this study. In bounce calculation, the accuracy and precision of trajectory were the main factors that influence the calculation effect. Therefore, we optimized the two-stage degree learning networks to provide the trained model with high accuracy and precision coordinates to facilitate the application of bounce detection and subsequent strategic analysis.