Fight detection by using motion co-occurrence feature

Fight detection is still an open problem for video concept detection. There has been little success in creating an algorithm that can detect fight in surveillance videos, movies, or TV broadcasts with high performance values. In this paper, we propose a new method for this task. Firstly, motion vectors are extracted by using optical flow block matching algorithm for the consecutive frames. Secondly, motion vectors are quantized with respect to their angles and magnitudes. Afterwards, velocity and angle based motion co-occurrence features are calculated by considering both current and past motion vectors. Experimental results obtained using kNN classifiers showed that the proposed algorithm can discriminate fighting scenes with high true positive rates and low false alarms.