AI-Powered Traffic Rule Violation Detection and Evidence Generation: A Multi-Model Deep Learning Framework for Urban Roads

Road accidents happen a lot in urban and developing areas because of improper driving behavior, such as riding without a helmet, lane misuse and unusual lane crossing. Law enforcement in these areas still depends on manual CCTV monitoring and there is very little high technology support. This research addresses that problem by implementing an AI-powered traffic rule violation detection and evidence generation framework, which integrates multiple models in a single pipeline so that several traffic cases can be detected under different environmental conditions. The system uses YOLOv8 for the detection of helmets, vehicles and number plates, PSPNet and SegFormer-B3 for the segmentation of lanes and road markings, and Faster R-CNN for reasoning about complex rule violations. The outputs of every model are combined through one inference pipeline that produces annotated images, videos and JSON evidence files. The dataset was collected from local traffic footage and public sources, and it was enhanced through augmentation, pseudo-labelling and a confidence-weighted lane probability formulation. Results show that YOLOv8 performs robustly, with mAP₅₀ = 0.93 for helmet detection and mAP₅₀ = 0.91 for number plate localization, while PSPNet achieves mIoU = 0.87 for lane segmentation and Faster R-CNN obtains precision = 0.91 and recall = 0.86 for rule violation classification. On multi-condition test data the overall system reaches 87.4 % violation detection accuracy. Bringing all the trained models into one framework provides an end to end solution for smart city enforcement and traffic analytics. Further tuning is still required for optimized real time deployment, but this study proves that a multi-model deep learning approach for this purpose is feasible and viable.