Real-time theft detection in urban surveillance: A comparative analysis of YOLO-based approach

The escalating concerns over urban security have led to the widespread deployment of camera surveillance systems. However, the reliance on human intervention limits their effectiveness and makes them vulnerable to security threats. To address this challenge, we present a novel automated system for real-time theft detection in urban environments using advanced YOLO deep learning models. The proposed system uniquely focuses on detecting anomalous or suspicious activities, such as vehicles with obscured license plates and unusual transport like horse-drawn carriages, which are indicative of theft-related incidents in urban settings. Our framework integrates real-time object detection with a license plate recognition module and a web-based alert interface, enabling multi-modal analysis and immediate response. We evaluate three YOLO model variants (YOLOv4, YOLOv5, YOLOv8) on a diverse surveillance dataset. Notably, YOLOv8 achieves detection accuracies of 98% for vehicles with hidden license plates, 99% for those with visible plates, and 97% for horse-drawn carts, outperforming the earlier YOLO versions in both precision and speed. Through a comprehensive comparative analysis, we demonstrate the superiority of the YOLOv8-based approach in balancing high accuracy and low inference time for real-time theft anomaly detection. The proposed system facilitates timely intervention by sending instant alerts and logging detailed data for each incident, thereby significantly strengthening urban surveillance capabilities. This work advances the state-of-the-art in intelligent security systems by introducing a robust and scalable framework for autonomous theft detection within real-world surveillance networks.