Event recognition in broadcast soccer videos

Abstract

Automatic recognition of important events in soccer broadcast videos plays a vital role in many applications including video summarization, indexing, content-based search, and in performance analysis of players and teams. This paper proposes an approach for soccer event recognition using deep convolutional features combined with domain-specific cues. For deep representation, we use the recently proposed trajectory based deep convolutional descriptor (TDD) [1] which samples and pools the discriminatively trained convolutional features around the improved trajectories. We further improve the performance by incorporating domain-specific knowledge based on camera view type and its position. The camera position and view type captures the statistics of occurrence of events in different play-field regions and zoom-level respectively. We conduct extensive experiments on 6 hour long soccer matches and show the effectiveness of deep video representation for soccer and the improvements obtained using domain-specific cues.