LHAT-YOLO: Study on intelligent monitoring algorithm for helmets at construction sites

TL;DR

This paper introduces LHAT-YOLO, a lightweight deep learning model for real-time helmet detection at construction sites to improve worker safety.

Contribution

The novel contribution is the development of LHAT-YOLO, a lightweight and accurate model for helmet detection using GSConv and FCD modules.

Findings

LHAT-YOLO reduces model complexity by 11% GFLOPs and 9.5% Params compared to YOLOv11.

The model achieves high accuracy with a Precision of 93.95% and Recall of 88.99%.

It maintains strong performance with an mAP50 of 94.92% on a dataset of construction site images.

Abstract

Construction sites in civil engineering projects are prone to sudden accidents. In particular, head injuries pose a significant safety threat to construction workers. Helmets play a vital role in protecting the heads of construction workers. Most construction sites still rely on manual methods to monitor workers’ helmet compliance, which is not only inefficient but also incapable of real-time monitoring. While traditional models can achieve intelligent monitoring, their lack of real-time capability fails to meet practical demands. However, the introduction of deep learning has transformed this situation. Therefore, this study proposes an intelligent monitoring method for helmet wearing at civil engineering construction sites based on deep learning theory. The study uses GSConv to improve the convolution module of the YOLOv11 deep learning model and adds a lightweight detection head: FCD (Fast Convolutional Detection) to establish the LHAT-YOLO model (Lightweight YOLO model for detecting helmets). While reducing the complexity of the model, it maintains accuracy and achieves efficient and intelligent detection of helmet wearing at construction sites. Experimental results show that on the dataset comprising 19,780 images in the training set, 2,473 images in the validation set, and 2,473 images in the test set, the LHAT-YOLO model reduces GFLOPs by 11% and Params by 9.5% compared to the YOLOv11 model. Overall, the LHAT-YOLO model achieves a Precision value of 93.95%, a Recall value of 88.99%, an mAP50 value of 94.92%, and an mAP50-95 value of 65.28%. This demonstrates that the LHAT-YOLO model maintains high accuracy even while being lightweight.

Figures