Rider surveillance to ensure wearing of helmet using deep learning

Abstract

This research paper focuses on using deep learning algorithms to automatically detect and monitor whether riders are wearing helmets properly. The proposed system employs a YOLOv5-based object detection model trained on a custom dataset of 12,000 annotated images, achieving 94.7% mAP in real-time traffic surveillance scenarios to assist law enforcement in ensuring road safety compliance.

Topics

Full Text Preview

We trained YOLOv5s and YOLOv5m variants on 12,000 images collected from traffic cameras at 15 intersections in Delhi. The dataset was annotated with four classes: helmet-worn, no-helmet, motorcycle, and rider. Data augmentation included random brightness, mosaic, and mixup. The best model (YOLOv5m) achieves 94.7% mAP@0.5 at 42 FPS on an NVIDIA RTX 3060, suitable for real-time deployment.

Published Through