Extracting 3D information from broadcast soccer video

In this paper, we propose a new method to estimate players' and ball's positions from monocular broadcast soccer video. With the relationship between objects and the camera in perspective projection, we derive the formula for estimating the moving objects' positions in real world, even when the ball is in the air. This method calibrates the camera's position in the stadium through the homography between the image and the playfield, and the self-calibration for rotating and zooming camera. Thus, the method can estimate the ball's position in the air without referring to other reference object with known height. In order to reduce manual interference, the players are detected based on the playfield detection. For the ball, we combine the detection procedure and tracking procedure organically. First, we extract candidate regions in each frame, then search the most likely regions in consecutive frames using Viterbi decoding algorithm. Once detected, the ball will be tracked by Kalman filter, which can help improve the detection recall. The system checks whether the ball is lost automatically. If it is lost, the detection procedure restarts. Experiments on synthesized data verify the proposed method, and promising results are obtained on real video data.