Classification of multi-agent trajectories
Jean-Charles Bricola · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · 2013
The aim of this project consists of identifying synchronized team activities with a particular focus on the recognition of basketball half-court offenses. We introduce a framework that uses the 2d-trajectory of each player to determine the class which the basketball tactic falls in. Every trajectory is segmented into tracklets which are then matched to some codewords. The occurence of the chosen codewords is used to build an histogram descriptor for each trajectory. Finally, the combination of all descriptors is used to recognize the activity. In this report, we determine the features and descriptors which are relevant for describing basketball activities. We then show that our framework achieves good recognition capabilities over different basketball datasets and that segments of activities can as well be successfully classified.