Deep learning-based real-time traffic rule violation detection system:

Abstract

Traffic rule violations are a significant issue in cities, causing bottlenecks and accidents due to slow-moving vehicles, rickshaws, and illegal vehicles. These infractions impede traffic flow and increase the risk of accidents. This research project proposes a real-time traffic rule violation detection system using the Yolo model to identify slow-moving vehicles or objects causing traffic jams or accidents. The system uses videos or images of vehicles or pedestrians in Bangladesh's capital city Dhaka as raw data. Three of the best object detection algorithms, YOLOv5, YOLOv7, and YOLOv8, were used for object detection. YoLov8 demonstrated the best accuracy, providing a single framework for training models for object identification, instance segmentation, and image classification. YOLOv8 took the crown for overall accuracy, with the highest mAP50 (0.94) and an excellent track record of accuracy in most classes. It is the most precise in both automobile and infraction detection, but its instance count is not as high as YoLov5. YoLov7, a well-balanced competitor, may match YoLov8's accuracy but can't match its respectable mAP50 and recall values. YoLov5, with the most detections (2097), takes the top spot, indicating a wider net but at the expense of some limitations