EdgeSafeAI: An AI-Based System for Real-Time Violence Detection

Surveillance systems have evolved dramatically in the last few decades due to advancements made possible by artificial intelligence (AI). For example, many current CCTV systems still utilize human operators to monitor video feeds; therefore there can be significant delays in response times and also many opportunities for human error to occur. Through the use of deep learning algorithms — such as the YOLO object detection model and MobileNetV2 for violence classification — this paper describes a virtual surveillance system built on AI that creates an ability for real-time detection of potential threats within the surveillance environment (e.g., weapons, violent acts, and fire hazards). The video frames are acquired and subsequently processed using OpenCV in order to detect clearly identifiable patterns within the video frame that indicate potential risk of danger to people or property (as defined by an organization). Once detected, the virtual surveillance system generates an automatic alert and sends a notification (with an attached snapshot) directly to the designated security officer(s). All detected incident data is stored in a centralized database, with access via a web-based interface, so that the military, law enforcement, and other agencies responsible for public safety can monitor incidents. Finally, the virtual surveillance system will enhance efficiency and reduce workload for security personnel; furthermore, it will allow agency personnel to respond more quickly as they will have full visibility of the controlled surroundings with current information. Overall, experimental results indicate that the virtual surveillance system is capable of detecting threats with high levels of reliability within a surveillance environment.