A Deep Learning Model for Helmet Detection and Automatic License Plate Recognition

Abstract

Not wearing a helmet among motorbike drivers is one of the leading causes of fatal accidents in developing countries. Detecting individuals not wearing helmets through license plates plays an important role in monitoring, reminding and punishing violators to help reduce accidents. Current models for detecting traffic violators through license plates are facing many limitations, such as difficulty detecting multiple vehicles in one frame, and ineffective methods for identifying license plates and people in the same vehicle and cannot simultaneously perform both parts: detecting people not wearing helmets and license plate recognition. Our proposed pipeline consists of two steps. Step 1 is to identify violating vehicles using YOLOv8m. Step 2 is to extract license plate information using the PaddleOCR library. To resolve confusion between objects, we use Bytetrack in first step to track and analyze. In both steps, post-processing techniques are developed to avoid errors in the identification process. This technique will take the class that appears most frequently in the motorcyclist's box and assign them together. Through testing, the model achieved high accuracy with mAP (mean Average Precision) is 97.9% and an accurate license plate recognition rate of 90.4%. Research results show that the proposed model achieves impressive efficiency in both tasks, helping to improve traffic safety and traffic management effectively