Abstract:

To address the issues of complex environmental interference and low recognition rates in early-stage tunnel fire detection, an improved YOLOX-based detection method, YOLOX-T, was proposed. The proposed method incorporated a NAM into the YOLOX network to suppress environmental noise and interference, thereby enhancing the model's robustness. A weighted BiFPN was integrated to improve multi-scale feature extraction and fusion. Furthermore, an α-IoU(Intersection over Union) loss function was employed to enhance the detection accuracy of early-stage tunnel smoke and flames, which often exhibit indistinct contours. Addressing the scarcity of publicly available datasets, a tunnel fire dataset encompassing both real-world and simulated scenarios was constructed through web data acquisition, simulated fire experiments, and the augmentation of existing datasets. Experimental results on the self-built dataset demonstrate that, compared to the original YOLOX model, the YOLOX-T method achieves improvements of 1.89% in mean Average Precision (mAP@0.5), 0.88% in mAP@0.5~0.95, 4.57% in precision, and 5.45% in recall. The improved algorithm can achieve better detection performance.

Key words: tunnel fire, YOLOX, normalization-based attention module (NAM), bidirectional feature pyramid network (BiFPN), fire detection