AI-Powered Construction Safety Monitoring Using Chain-of-Thought Verification

A new research paper proposes a passive, end-of-shift construction safety monitoring system that uses a three-stage AI pipeline to detect hazards and verify compliance. The system processes video from body-worn and fixed cameras through fine-tuned YOLO11 for PPE and hazard detection, SAM 3 for segmentation and worker deduplication, and Qwen3-VL-8B-Instruct with a persona-scaffolded adversarial chain-of-thought protocol for verification. The key innovation is the Stage 3 prompt design, which uses professional persona backstories to achieve a 12% precision improvement over single-pass methods. The research addresses the 1,055 fatal worker injuries recorded in the US in 2023, aiming to reduce preventable deaths through automated monitoring.