Abstract

This paper proposes a novel ball tracking approach for coping with difficult situations as occlusion and fast object movement, in the context of collective sports. In particular, in the context of soccer, the ball cannot be represented by the features which are commonly utilised in the state of the art, because of the high deformation of the ball in case of fast movement, and considering the small size of the ball (below 30 pixels), due to the distance of the static cameras with respect to football ground. In this work, we propose an elliptical model able to represent the deformation of the ball when is moving fast, which combines elliptical shape features with its current appearance, and the available information about the object dynamics. For ball tracking, we apply a two stage algorithm. At the first stage, a ball image matching method is applied, finding all the candidate regions for the position of the ball in the field. At the second stage, the detected regions are analysed in relation with the dynamics of the previously detected ball candidates, and their similarity with regions in the previous frames, rejecting candidates when the current region is more similar to previous no-ball regions in comparison to previous ball regions. If the ball is occluded by players or camouflaged by field lines, we apply a new additional method for recovering the position of the ball. The main contributions of our method are: the incorporation of the ball deformation to the ball model, and a new algorithm for the detection of the ball in occlusion situations. We have tested our approach in real benchmark sequences, facing complex challenges related to occlusion, low resolution, deformation due to speed, and illumination problems, with very competitive results.