Generative AI-driven data augmentation for enhanced construction hazard detection

Abstract

The construction industry has long struggled with poor safety records. Traditional safety monitoring methods, reliant on manual observations, are often ineffective. To address these limitations, computer vision and generative artificial intelligence (AI) have been explored. While computer vision has shown promise in automating safety monitoring, its effectiveness is often hindered by the challenges of efficiently collecting diverse datasets. Generative AI offers a potential solution by augmenting image datasets, enabling more robust construction hazard detection. This paper investigates the use of generative AI for augmenting image data to improve hazard detection performance. Various combinations of generative AI tools and prompting strategies are tested. The results show that the combination of image-guided structured prompting with Stable Diffusion achieves the highest detection performance (mAP@50 of 92.5 %) using 150 augmented images. This represents a substantial improvement compared to the baseline mAP@50 of 51.6 % achieved with real images alone.