Action scene detection from motion and events

Action scenes usually contain higher motion activity than other scenes in feature films while showing events like fights, gun shots and car crashes. This work investigates motion and event detection to separate action scenes from non-action scenes. In contrast to existing work, the proposed system does not consider the shot structure of video. The approach uses SVMs to classify GIST-based global motion features, SIFT-based local motion features, and bag of MPEG-7 ColorLayout features. Two test sets of Hollywood movies and user generated action movies are used to evaluate the system. The results of a frame-level evaluation indicate that especially, the global motion approach represents a good tradeoff between accuracy and speed. A scene-level evaluation shows that the combined system compares well to existing shot-based approaches.