A Real-Time Framework for Anomaly Detection in CCTV Surveillance Systems

Abstract

Modern surveillance solutions are increasingly leveraging intelligent video analytics to enhance public safety, especially in high-risk areas. This project proposes a dual-model deep learning system that detects both firearms and violent behavior in real time. The first component, a firearm detection module, uses the YOLO algorithm—famously processing images at speeds up to 45 FPS, with newer versions like YOLOv8 achieving ~85 % precision and 0.8 s/frame real-time performance. The second component, a violence recognition module, relies on a Vision Transformer (ViT-B/32), which excels at learning spatial and temporal patterns in video sequences. Transformer-based models such as ViViT have demonstrated strong performance in identifying violent actions like fights and shoves. By combining these systems, the surveillance setup triggers alerts only when both weapons and aggressive behavior are detected—significantly reducing false alarms and enhancing situational awareness. This integrated approach is ideally suited for deployment in schools, transit hubs, and other sensitive public spaces, enabling timely and reliable crime prevention.