Automated Helmet Detection in Safety Work Environments Using YOLOv5

Abstract

Ensuring workplace safety in an in-plant workplace or industrial environment is very critical, where the use of protective gear, particularly helmets, is crucial in preventing severe injury. Even if strict measures have been introduced, non-compliance remains a significant issue due to human negligence and a lack of continuous monitoring. This research aims to develop an automated, real-time helmet detection system leveraging computer vision to enforce safety regulations and minimise accidents related to the absence of helmets. The proposed solution uses YOLOv5 (You Only Look Once version 5), a state-of-theart object detection algorithm that is best known for its high accuracy and real-time performance. The model was trained on a publicly available helmet detection dataset from Kaggle, comprising 764 images labelled with bounding boxes in PASCAL VOC format, categorised into two classes: With Helmet and Without Helmet. To address dataset limitations, data augmentation and transfer learning were employed. Training and evaluation were conducted using PyTorch on a GPU-enabled environment. Performance metrics included precision, recall, mAP (mean Average Precision), and real-time inference speed. The trained YOLOv5 model demonstrated robust performance in distinguishing between helmeted and non-helmeted individuals, achieving a high mAP and real-time inference speed suitable for deployment in surveillance systems in safety-critical zones. The developed detection system shows promise as an effective tool for enforcing helmet compliance in safety-critical environments. Future work may include multi-class detection (e.g., detecting safety vests), integration with CCTV systems, and deployment of edge devices for real-time alerts.