A monocular vision-based method using DL-YOLO for fire detection and localization in road tunnels

Vision-based fire detection and localization in road tunnels remain challenging due to complex illumination, multi-scale fire targets, the lack of realistic tunnel fire datasets, and the unsolved problem of smoke backflow front positioning. To address these issues, we propose DL‑YOLO, a unified framework integrating an enhanced YOLOv8n detector with a monocular geometric localization module. The detector incorporates SPD‑Conv to preserve fine‑grained details under low contrast, a global attention mechanism to suppress tunnel‑specific interferences, an auxiliary P2 head to improve small‑target recall, and Ghost modules for computational efficiency. A full‑scale tunnel fire dataset was constructed through experiments in six tunnels, and a relative‑scale criterion is proposed to define small targets adaptively across varying resolutions. The detector achieves 96.9% mAP@0.5 and 93.6% recall, outperforming the baseline by 2.1% and 4.7%, respectively. The localization module leverages tunnel structural priors and camera parameters to estimate spatial coordinates of both flame sources and smoke backflow fronts, achieving a mean absolute error of 0.51 m (1.55% relative) for smoke‑front positioning, and 1.00 m (1.80% relative) in longitudinal error and 0.12 m (2.7% relative) in lateral error for flame sources. Sensitivity analysis confirms that localization errors propagate predictably within practical uncertainty bounds. Both field experiments and controlled tunnel tests validate the method’s effectiveness for real‑time tunnel fire monitoring.