Discovering tactics in broadcast sports video with trajectories

Abstract

Unlike traditional audience oriented semantic analysis in sports video, tactic analysis aims to assist the professionals to improve their training performance and adopt proper tactics in the game. Trajectories of ball and players contain temporal and spatial motion information, which is a representative feature for tactic analysis. We propose a generic scheme to discover tactics in classified trajectories in the same events. Local similarity measurement between trajectories is developed for trajectory matching to reduce the noise in the broadcast sports video. Transductive classifier with the learnt bias of low-level features of the trajectories is introduced to deal with different games and classify trajectories into tactic events only using a small labeled dataset. Frequent pattern mining is employed to find the tactics based on the subsequences of the classified trajectories. Experiments show that the proposed approach achieves promising tactics discovery results in different broadcast sports game videos.