Real-time Violence Activity Detection Using Deep Neural Networks in a CCTV camera

In surveillance systems one of the most difficult and critical tasks is detecting violence, most video surveillance systems are faced with the challenges of false alarms and working in a real-time environment. This paper introduces a lightweight system that can be employed in real-time systems. The system proposed uses OpenPose for multi-person 2D pose estimation, YoloV3 for person detection, and a CNN to classify the violent action. OpenPose unlike other systems uses a bottom-up approach which decouples the running time from the number of people in the frame, YoloV3 deployed for person detection has a very fast processing time and the CNN used was trained on the skeleton dataset that was generated and had very good accuracy. We also propose an skeleton image dataset of three action categories, namely kicking, punching, and non-violent.