Automatic fight detection in surveillance videos

Affective computing is an up-surging research area relying on multimodal multimedia information processing techniques to study human interaction. Social signal processing under affective computing aims at recognizing and extracting useful human social interaction patterns. Fight is a common social interaction in real life. A fight detection system finds wide applications, such as in a prison, a bar and so on. Research works on fight detection are often based on visual features and demand substantive computation and good video quality. In this paper, we propose an approach to detect fights in a natural and low cost manner through motion analysis. Most existing works evaluated their algorithms on public datasets manifesting simulated fights, where the fight events are acted by actors. To evaluate on real fight scenarios, we collect fight videos from YouTube to form our own dataset. Based on the two types of datasets, we process the motion information to achieve fight detection. Experimental results indicate that our approach accurately detect fights in real scenarios. More importantly, we uncover some fundamental differences between real and simulated fights and could discriminate them well.