Spatio-temporal Classification of Aggression in Video Surveillance using Optical Flow History

Abstract

The role of automatic surveillance in modern society is rapidly in- creasing. While most of these systems operate by making a judgment of aggression based on two frames, in this paper we explore the effect of adding more frames to the quality of detecting aggression in video surveil- lance. The premise of doing this is that the system must be able to run in real-time, preferably using as little computing power as possible. We evaluate an algorithm by TNO that was developed with this goal in mind. The results show a significant increase in detecting aggressive instances when more frames are added. The tests were run on a dataset containing instances of street aggression supplied by the Dutch police. A history of 1.5 seconds worth of frames was found to deliver the best results on the dataset.