Feature-based Localization Refinement of Players in Soccer using Plausibility Maps

In soccer, live acquisition of quantitative motion information such as distances covered by players or ball possession can only be performed by sophisticated automatic algorithms. The most challenging task in automatic acquisition of player movements arises in situations when players occlude each other. The main contribution of this work is a computer vision based localization of individual players in real-time with focus on crowded situations. The system automatically detects occlusions between players and performs an advanced refinement of the player localization. This is done by creating a plausibility map that assigns to each image pixel a confidence of being a player's center point. The plausibility map is based on the players' heights, the colors of their jerseys and the densities of their visual appearances in the image. Results confirm a significant improvement of the performance of tracking players in soccer scenarios.