EVALUATION OF DEEP LEARNING MODELS FOR ACCURATE SAFETY HELMET DETECTION IN CONSTRUCTION ENVIRONMENTS

Abstract

Abstract— The safety of construction workers can be significantly enhanced by wearing safety helmets. Workers frequently remove their helmets due to a lack of knowledge and discomfort, exposing them to hidden dangers. Accidents involving falling objects from heights put workers who are not wearing safety helmets at increased risk. Therefore, there is a critical need for a fast and accurate safety helmet detector to oversee construction site safety. However, standard manual monitoring is time-consuming and labor-intensive, and solutions that include attaching sensors to helmets are not generally used. To address this issue, this study presents a Deep Learning (DL)--based technique for accurately and rapidly detecting safety helmets. The data for training the DL models is sourced from the Kaggle site and goes through preprocessing stages such as image improvement and resizing. Helmet detection utilizes DL models such as You Only Look Once (YOLO), Single Shot MultiBox Detector (SSD), and Region-based Convolutional Neural Network (RCNN). The experimental results show that the YOLO model has the highest accuracy, precision, recall, and mean average precision (mAP), at 94.7%, 93.8%, 94.2%, and 94.62%, respectively. These findings highlight the effectiveness of YOLO in real-time helmet identification, greatly contributing to accident avoidance on construction sites.