AI-Powered Traffic Violation Detection System Using CCTV, and Deep Learning

Abstract—Urban road safety continues to deteriorate alongside growing traffic density, with traffic violations remaining a leading causal factor in accident-related deaths. Traditional enforcement approaches relying on manual surveillance and passive CCTV recording are systematically inadequate given the scale, speed, and complexity of contemporary traffic environments. This paper presents the design, implementation, and empirical analysis of an AI-Powered Traffic Violation Detection System that augments existing CCTV infrastructure with deep learning inference pipelines to enable fully automated, real-time violation identification and evidence generation. The proposed system employs YOLOv8 as its anchor-free, single-stage object detector, enabling simultaneous detection of five violation types: helmetless riding, traffic light jumping, triple riding, seatbelt non-compliance, and wrong-way driving. Multi-object tracking is achieved through DeepSORT integration, which ensures temporal identity continuity. Automatic License Plate Recognition (ALPR) is performed via a YOLOv8-based plate localization model combined with EasyOCR character recognition. Violation evidence is persistently stored in an SQLite database, while a Streamlit dashboard provides both real-time monitoring and historical analytics. The system achieves a mean Average Precision (mAP@0.5) of 94.0% on a test set of captured real-time frames, a throughput of 28 FPS on GPU, whole-plate recognition accuracy of 88.4%, and evidence logging latency under 180 ms, validating its feasibility for smart traffic policing. Index Terms—Traffic Violation Detection, YOLOv8, Deep Learning, Computer Vision, DeepSORT, Multi-Object Tracking, ALPR, EasyOCR, CCTV Surveillance, Image capturing, Real-Time Video Processing.