Enhancing Worker Safety at Heights: A Deep Learning Model for Detecting Helmets and Harnesses Using DETR Architecture

Abstract

We propose an approach for monitoring work-at-heights activities by ensuring the presence of safety helmets and harnesses in construction environments using the DEtection TRansformer (DETR) model. The DETR-based model is trained and tested on a custom dataset to accurately detect these essential safety items. The model’s performance was evaluated using standard metrics. For helmet detection, it achieved an average precision of 0.99, a recall of 0.92, an F1 score of 0.95, and a confidence rate exceeding 95%. Similarly, for harness detection, the model attained an average precision of 0.971, a recall of 0.91, an F1 score of 0.94, and a confidence rate above 93%. To evaluate its reliability, the model was tested under diverse conditions, including RGB, grayscale, blurred, and dusty images, as well as scenarios involving multiple helmets and harnesses. It was also assessed using drone-captured images to ensure robust performance across varying weather and lighting conditions. The results were then compared with a previously trained YOLOv3 model developed for the same detection task, as well as two newly trained models based on the YOLOv8 and Deformable DETR architectures. Our analysis demonstrates that the DETR model not only improves detection accuracy but also overcomes certain limitations inherent in other models. The comparative study highlights key architectural and performance differences, providing insights into the advantages of transformer-based models for object detection tasks in industrial safety applications.