Fight Detection in Images Using Postural Analysis

This paper defines the research process that has been carried out to develop a system for detecting fights in images. The system takes input frames and evaluates the probability that the frame contains one or more people fighting. Input frames containing images are initially processed using a well-known neural architecture, called OpenPose, which extracts pose information out of images that contain human postures. Human posture data is then processed by heuristics to extract both: angles for arms and legs for each person in the image. Angles are then used to feed an additional neural network that has been trained to make probability predictions of people being potentially involved in fights. This paper describes the full pipeline regarding techniques, tools and assessment required to create a camera based violence detection system.