Audio-visual football video analysis, from structure detection to attention analysis

Sport video is an important video genre. Content-based sports video analysis attracts great interest from both industry and academic fields. A sports video is characterised by repetitive temporal structures, relatively plain contents, and strong spatio-temporal variations, such as quick camera switches and swift local motions. It is necessary to develop specific techniques for content-based sports video analysis to utilise these characteristics. For an efficient and effective sports video analysis system, there are three fundamental questions: (1) what are key stories for sports videos; (2) what incurs viewer’s interest; and (3) how to identify game highlights. This thesis is developed around these questions. We approached these questions from two different perspectives and in turn three research contributions are presented, namely, replay detection, attack temporal structure decomposition, and attention-based highlight identification. Replay segments convey the most important contents in sports videos. It is an efficient approach to collect game highlights by detecting replay segments. However, replay is an artefact of editing, which improves with advances in video editing tools. The composition of replay is complex, which includes logo transitions, slow motions, viewpoint switches and normal speed video clips. Since logo transition clips are pervasive in game collections of FIFA World Cup 2002, FIFA World Cup 2006 and UEFA Championship 2006, we take logo transition detection as an effective replacement of replay detection. A two-pass system was developed, including a five-layer adaboost classifier and a logo template matching throughout an entire video. The five-layer adaboost utilises shot duration, average game pitch ratio, average motion, sequential colour histogram and shot frequency between two neighbouring logo transitions, to filter out logo transition candidates. Subsequently, a logo template is constructed and employed to find all transition logo sequences. The precision and recall of this system in replay detection is 100% in a five-game evaluation collection.