New content-based features for the distinction of violent videos and martial arts

Real violence is unwanted content in video portals as it is forensically relevant in video surveillance systems. Naturally, both domains have to deal with mass data which makes the detection of violence by hand an impossible task. We introduce one component of a system for automated violence detection from video content: the differentiation of real violence and martial arts videos. In particular, we introduce two new feature transformations for jitter detection and local interest point detection with Gestalt laws. Descriptions are classified in a two-step machine learning process. The experimental results are highly encouraging: the novel features perform exceptionally well and the classification process delivers practically acceptable recall and precision.