Vision-language model-based intelligent assistant for onsite construction safety inspection

초록

Construction site inspection demands a contextual understanding of dynamic job-site conditions, traditionally relying on inspectors' expertise combined with adherence to predefined safety regulations and industry standards to identify hazards. While vision-language models can detect and describe hazards, they struggle to correlate observations with regulations due to limitations in geometric reasoning. Recent studies show progress in compliance checking, but these models are still challenged by dynamic scenarios. This paper introduces a hybrid framework that combines geometric reasoning with LLM-powered interpretation to improve regulationaware hazard detection. The system achieved high detection accuracy: 97 % for ladder use, 94.6 % for mobile scaffolding, and 99 % for fire-related work. Captioning performance evaluated through BLEU, ROUGE, METEOR, and BERT Score showed strong semantic alignment. User feedback confirmed its efficiency and ease of use, even under dynamic conditions. By integrating visual data with regulatory reasoning, the proposed system offers a practical, domain-adapted solution for enhancing construction safety inspections.

키워드