Dominant interaction group detection in team sports

With sport performance analysis shifting towards a team’s overall behaviour, the need for automatically discovering formations, such as players temporarily forming a common group of interactions, emerges to complement the coach’s observations. We propose a novel framework to detect a structured group of players, encoded in the form of context-dependent team players’ interactions. In real scenarios, considering a tendency among the players’ movements to occupy a common interested region on a sport field, we predict the future candidate area of group interactions (tendentious zone) before the group formations occur. Consequently, the tendentious zone guides the future players’ movements and provides prior information about their future positions on the field. Building a graph of all players’ positions and considering their motion stability towards the tendentious zone, we aim to discover an optimal subgraph indicating a dominant group of players by maximising the similarity among them. To quantify the similarity of any two players, we
consider their relative proximity as well as the common social attention model. Experiments on new sports datasets consistently show the superiority and effectiveness of the proposed approach over existing group detection methods.