Real-time multi-view event detection in soccer games

In the last decade, several research efforts have been undertaken in soccer video analysis. This increasing interest is motivated by the possible applications over a wide spectrum of topics: indexing, summarization, video enhancement, team and players statistics, tactic analysis, refereepsilas support, etc. Soccer video analysis requires different challenging tasks: ball and players have to be localized in each frame, tracked over time and, above all, their interactions (passes/shoots) have to be detected and analyzed. The latter task is fundamental, especially for statistics and referee decision support purposes, but, unfortunately, it has not received adequate attention from the scientific community. In this paper a multi-view system able to understand in real time the interactions between the ball and the players is presented. The 3D ball trajectories are, firstly, extracted by triangulation from multiple fixed cameras and then projected on a virtual play-field where they are temporally analyzed to detect their variations generated by the interaction with the players. Inference processes are then introduced to fix the instant of the detected interaction and, finally, the player kicking the ball is identified by analyzing human body configuration with an innovative neural approach based on a Contourlet representation of human silhouette data. The system has been tested during several matches of the Italian first division football championship and experimental proofs of its effectiveness are reported.