Spatio Temporal Feature-based Weapon and Violence Detection from Video Surveillance using Deep Learning

The increasing incidence of violence in public spaces has made developing efficient and accurate weapon detection systems necessary. A Long Short-Term Memory (LSTM) model built on the MobileNet structure has been proposed for the detection of violence and weapons. It examines each frame of a video stream to look for indications of violence or the presence of weapons. The proposed system is trained on a dataset of simulated violent incidents. The dataset includes fights, gunfights, knife attacks, and diverse environmental and lighting conditions. The system is excellent for use in surveillance and security systems and 93.86% accuracy is achieved. It can be implemented on embedded systems or mobile devices due to its high efficiency and minimal computing demands. The proposed method can increase public safety by notably increasing the efficiency of security and surveillance systems.