Robust flame detection in complex tunnel fire environments: A scaled experimental study using SDS-YOLO

Owing to the enclosed structure and complex environment of highway tunnels, tunnel fires pose a growing threat to traffic safety and emergency evacuation. Existing object detection frameworks frequently struggle to balance detection performance with high-fidelity identification of flame features under smoke obscuration and vehicular illumination interference. An enhanced tunnel fire detection model, the SDS-YOLO algorithm, is developed based on the YOLOv11 baseline framework to address these limitations. By integrating SPD-Conv module, DySample dynamic upsampling operator, and SEAM attention mechanism, this model is engineered to heighten perceptual robustness toward flame signatures under low-visibility conditions and intense background lighting noise. Concurrently, a comprehensive tunnel fire dataset is constructed utilizing a 1:20 scaled physical experimental platform, covering multiple fuel combustion types, different smoke concentration gradients, and complex illumination interference conditions. Experimental results demonstrate that SDS-YOLO achieves mAP50 values of 87.2% on the multi-level smoke dataset and 86.7% on the complex illumination interference dataset, while maintaining only 4.1 M parameters and an inference speed of 112 FPS. Under extreme conditions involving dense smoke obscuration and severe glare interference, SDS-YOLO model significantly outperforms the advanced models of the YOLO series and DETR, demonstrating superior anti-interference capability and strong potential for industrial-grade real-time tunnel fire monitoring.