Real-time Violence Detection using Deep Learning Techniques

The subject of violence detection plays a significant role in tackling threats and abuses in society. It is the key element of any security enforcing system. The widespread deployment of video surveillance has facilitated the law enforcement agencies to visually monitor environments and take prompt action in case of any alerting situation. This task requires manual interaction for continuously overseeing the live streams of CCTVs. This paper presents an efficient approach for detecting violence in real-time using different deep learning methods which diminishes the element of human supervision to a higher extent. The existing research on the topic of violence detection using machine learning is either based on specially created videos or immensely relies upon less accurate algorithms and infeasible assumptions. The presented system in the paper is premised on a hybrid approach of employing different algorithms for assessing all distinct aspects of the problem in a viable and effective manner. The proposed system is reliant upon YOLO for real-time object detection and Long Short-Term Memory for developing the classification module. DeepSort algorithm in the proposed approach further augments the efficiency. The model was trained using a relevant violence detection dataset and integrated with different software frameworks for enhancing the interface. As an outcome of the paper, we developed a fully-fledged violence detection system based on deep learning algorithms which passed different tests and evaluations.