Probabilistic tracking of the soccer ball

초록

This paper proposes an algorithm for tracking the ball in a soccer video sequence. Two major issues in ball tracking are 1) the image portion of the ball in a frame is very small, having blurred white color, and 2) the interaction with players causes overlapping or occlusion and makes it almost impossible to detect the ball area in a frame or consecutive frames. The first is solved by accumulating the image measurements in time after removing the players' blobs. The resultant image plays a role of proposal density for generating random particles in particle filtering. The second problem makes the ball invisible for time periods. Our tracker then finds adjacent players, marks them as potential ball holders, and pursues them until a new accumulated measurement sufficient for the ball tracking comes out. The experiment shows a good performance on a pretty long soccer match sequence in spite of the ball being frequently occluded by players.