Video analysis for segmentation and classification of players at soccer games

This paper presents a methodology for the segmentation and classification of players in videos of soccer games, which consist in three stages: segmentation, feature extraction and classification. In our methodology, segmentation is carried out using a Mixture of Gaussians Model, MOG, that allows to determine those regions with higher probability of belonging to a player. The number of probable regions is reduced by imposing morphological constrains. Characterization of the most probable regions are performed with a color descriptor in the RGB space, which is a 3D histogram. For classification, a Bayesian estimator is employed using the Bhattacharyya distance as likelihood function. Experiments shows that the proposed method works well with the same videos used for TV broadcasting. MOG and the RGB 3D histogram are robust and computationaly light in this case since the most distinguishable feature among players is the color of their jerseys. Performance evaluation was obtained comparing the results of the method against a human-labeled database.