Player tracking and identification of game systems in basketball using three cameras

In this paper a new strategy for observing and analyzing a basketball match using video processing techniques to identify the game systems of a team is described. The system tracks players' positions during the match. At the outset, three video streams from three fixed cameras are available, each processed separately to deliver measures of the players' positions from different available views. Each treated view includes foreground detection and a bounding-box tracker designed to contain the pixels representing each player. During the multi-view process, measurements from different views are synchronized to enable identification of the same player when the player is visible simultaneously on several cameras. These measurements are combined in order to obtain the players' positions as well as their updated positions through the images. The position thus obtained is exploited in a database containing the representative points (coordinates) of all the players, who form a polygon. The analysis of a game system is thus simply the analysis of the deformation and movement of this polygon during the match. Comparative indicators of the two teams are defined, along with an indicator which represents the opinion of an expert in the sport (action code). Statistical tools are exploited with the objective on the one hand of identifying correlations and relationships between different indicators, and on the other hand of identifying the game system adopted by comparing the expert opinion and the results of the established heuristic models.