Badminton Shuttlecock Tracking and 3D Trajectory Estimation From Video

Abstract

Being able to track the location of a shuttlecock in badminton videos can provide rich information for applications such as automated analysis of player tactics, prediction of the game progression, and content-based video retrieval. In this thesis, we analyze two main problems: where the shuttlecock is in each frame of the video (2-dimensional position tracking), and where the shuttlecock is in the 3D world based on the 2D trajectory of the shuttlecock (3-dimensional position estimation). We divide the 2D tracking problem into candidate selection, data association, trajectory classification, and trajectory linking. Using this method, we showed that it is possible to locate the shuttlecock in 85% to 92% of the frames with less than 10-pixel errors from the true position. We also attempt to reconstruct the 3D trajectories of the shuttlecock based on synthetic 2D trajectories. Using linear regression, SVM classification, and by solving a minimization problem, we showed that it is possible to achieve an error of 0.034 court-length (48 cm on a 13.4 meter court).