Violent behaviour detection using local trajectory response

Surveillance systems in the United Kingdom are prominent, and the number of installed cameras is estimated to be around 1.8 million. It is common for a single person to watch multiple live video feeds when conducting active surveillance, and past research has shown that a person's effectiveness at successfully identifying an event of interest diminishes the more monitors they must observe. We propose using computer vision techniques to produce a system that can accurately identify scenes of violent behaviour. In this paper we outline three measures of motion trajectory that when combined produce a response map that highlights regions within frames that contain behaviour typical of violence based on local information. Our proposed method demonstrates state-of-the-art classification ability when given the task of distinguishing between violent and non-violent behaviour across a wide variety of violent data, including real-world surveillance footage obtained from local police organisations.