Understanding Sport Activities from Correspondences of Clustered Trajectories

Human activity recognition is a fundamental problem in computer vision with many applications such as video retrieval, automatic visual surveillance and human computer interaction. Sports represent one of the most viewed content on digital tv and the web. Automatically collected statistics of team sports game play represent actionable information for many end users such as coaches and broadcast speakers. Many computer vision methods applied to sport activity classification are often based on multi-camera setups, player tracking and exploit information on the ground-plane. In this work we overcome this limitations and propose an approach that exploits the spatio-temporal structure of a video grouping local spatio-temporal features unsupervisedly. Our robust representation allows to measure video similarity making correspondences among arbitrary patterns. We tested our method on two dataset of Volleyball and Soccer actions outperforming previous results by a large margin. Finally we show how our representation allows to highlight discriminative regions for each action.