Keywords: construction safety, hazard prevention, pre-trained models, image recognition, RAG, LLM.

Previous studies on construction site hazard identification have primarily employed machine learning techniques to predict potential accident scenarios. However, two major limitations persist: (1) these models often rely on manual input of attribute parameters, reducing their applicability in real-time and automated settings, and (2) most approaches focus solely on hazard prediction without providing specific, regulation-compliant preventive strategies. To address these gaps, this study proposes an automated framework that integrates computer vision with Retrieval-Augmented Generation (RAG) for hazard identification and response plan generation. Specifically, hazard types are detected from construction site CCTV footage using computer vision, while large language models (LLMs) are employed to retrieve relevant construction safety regulations and generate corresponding mitigation measures. Empirical validation was conducted using a dataset of 2,490 hazard images to test the proposed model. Results demonstrate that the LLM-RAG framework can generate feasible, regulation-aligned preventive recommendations. The model significantly enhances the automation and intelligence of hazard recognition and mitigation planning, offering a novel approach to advancing smart construction safety management.

download the full-text of this paper (PDF, 12 pages, 563 Kb)

go to the previous paper
go to the next paper
return to the table of contents
return to the volume description