Deep Learning-Based Automatic Helmet Recognition for Two-Wheeled Road Safety

Abstract

Helmets are one of the most critical road safety precautions for two-wheeled participants (motorcycles, bicycles, and e-scooter) and not wearing protective helmets might cause severe or catastrophic injuries. The primary method for detecting helmets is currently a sequence of Convolutional Neural Network algorithms. To accomplish road safety, detection precision, forecast speed, and deployment simplicity are crucial factors. Traditional object detection techniques frequently fall short of achieving balanced effects in all domains. This research proposes a helmet detection application based on the most recent YOLOv7 algorithm with an attention-based improvement mechanism. The model's performance has been evaluated on a set of helmet test images with an average precision (mAP@0.5) reaching 91.4 %. As results indicate, high detection precision and low computational demands are achieved, making the model suitable for real-world deployment. Therefore, the proposed model can contribute to the solution of the problem involving helmet detection on two-wheeled vehicles.