Generating Ball Trajectory in Soccer Video Sequences

This paper demonstrates innovative techniques for estimating the trajectory of a soccer ball, using fixed cameras with constant calibration parameters. In contrast with broadcast coverage, for fixed camera data, the ball is often rendered with poor resolution away from the image centre. The rapidly moving ball is subject to motion-blur, caused by finite shutter speeds and interlaced fields, resulting in variable shape, size and colour. The velocity estimated from Kalman tracking is used in both normalising ball size and filtering the ball from false alarms. Furthermore, occlusion-reasoning and tracking-back methods are utilized to estimate its position when it is occluded, and also to remove false alarms. Finally, temporal hysteresis based thresholding of the ball likelihood is applied for trajectory filtering to improve the robustness and continuity of the tracked ball. Promising experimental results from several long sequences are reported.