The method for extracting player and ball coordinates using soccer tactical video

Abstract

Recently, a lot of tactic analysis has been done based on tracking data in soccer. Detecting the players and the ball is necessary to obtain tracking data using soccer video. However, detecting the ball information is problematic because the ball usually looks very small in a tactical video. This study proposes a method for extracting players and ball x and y coordinates: detecting balls in a tactical video, separating teams of detected players, and camera calibration. First, the proposed method uses the trained Faster-RCNN on the tactical video dataset to detect a ball. Multiple patch images are segmented by the image. Second, camera calibration work can extract the 2-dimensional position of players and balls. Third, we cluster detected players on the field into five clusters using the color information to extract the coordinates of each distinct cluster. An experiment is conducted to validate the proposed ball detection procedure, and the proposed method can improve ball detection performance. In addition, the camera calibration performance and clustering performance evaluation in the experiment are presented.