Abstract:
Road traffic crashes being a serious concern in Bangladesh, and other developing
countries, lack of helmets use and non-registered vehicles are significant causes of deaths
and injuries. Conventional traffic observation is manually conducted and labour-intensive
and can be error-prone and inadequate for enforcement at large scale. In this paper we
present helmet detection and automatic license plate recognition system based on machine
learning framework using YOLO models such as YOLOv8L, YOLOv12n and YOLOv12L.
We train the models on a large dataset which contains 6,176 images with corresponding
labels, mined from Kaggle that contains riders with helmet and no helmet wearing with
visible license plates in diverse environmental conditions. Augmentation approaches like
rotation, flip and contrast enhancement are used to further enhance real-life robustness
and generalization. The models are tested on major performance measures such as
precision, recall, mAP50, and inference time. With an inference speed of less than 5 ms per
image, YOLOv12L is orders-of-magnitude better performing than some solutions that
achieve 96% accuracy and 98% mAP50. Comparative experiments with CNN, ResNet,
YOLOv5 and YOLOv7 demonstrate that YOLOv12 outperforms others not only in the
accuracy but also efficiency. Also, Optical Character Recognition (OCR) support enables
remarkable recognition of the license plates with accuracy greater than 97% detected. For
this, we design a prototype system to automate helmet violations detection and license
plate details obtaining from surveillance videos to minimize the burden of manual
monitoring. This model provides scalable, effective, and adaptable applications for the
smart traffic management, and has a solid promise of implementation in mobile/web, and
supporting sustainable law enforcement and regulatory compliance.
