Suspicious behavior and object detection at traffic signals using YOLOv8

Traffic violations and security threats at intersections have been aggravated by the high rate of urbanization and vehicle congestion. Loitering, signal jumping, objects drop or roadside fights are some of the suspicious behaviors that are sometimes hard to monitor manually unless they are extreme or highly visible. In this paper, I describe a traffic vigilance and anomaly detection system powered by deep learning and utilizing the current state-of-the-art object detector, i.e., YOLOv8, combined with the features of real-time video analytics and multi-object tracking (DeepSORT). The system is highly trained on a curated dataset of publicly available data including AI City, CityFlow, UA-DETRAC, and UCF-Crime, and encompasses the broad set of traffic offenses and suspicious behaviors. Comparative experiments show that YOLOv8 performs better than the baseline models (YOLOv5 and SSD MobileNet), is more accurate (mAP of 90.5) and capable of inference in real time (32 FPS) with fewer false positives. In order to improve usability, live monitoring and alert generation via a Gradio-based interface are supported. Moreover, the effectiveness of the system is confirmed by a practical case study on traffic intersection footage, which proves the system to be strong in practice. The suggested solution adds to the infrastructural components of smart cities because it facilitates proactive policing, automated detecting inconsistencies, and improved citizen safety.