Full-scale sensing of moving fire hazards in tunnels using segmentation of external smoke images

Using tunnel external smoke to predict fires is an attractive method that helps firefighters rapidly grasp fire information. However, its application faces challenges from complex image background, moving vehicles, and cross-scale predictions. To solve these, a 1:100 scaled tunnel with moving fire sources is designed based on scaling laws. By varying fire power, velocity, and soot yield rate, a robust dataset of 27,000 smoke images is established. A Swin-Unet semantic segmentation model is proposed to extract smoke features under various tunnel backgrounds, after which a VGG-16 model is trained on the segmented images to predict fire power and location in real time. Results show that this segmentation-assisted method, using only front-view smoke images as input, achieves a superior balance between training cost and prediction accuracy (R 2 = 0.86, MSE = 0.17), with a fire velocity tolerance up to 5 m/s (actual speed of 180 km/h). Full-scale tunnel fires then validate the method and reveal the key to applying small-scale prediction model to a real tunnel: using a camera with fixed operational parameters, placed at a shooting distance following the proportional similarity law. With this setup, the model effectively blocks background interference and yields a location error below 9 m and an HRR error within 30% in an 80-m tunnel. This paper is expected to promote the intellectualization of tunnel fire safety systems at low installation and maintenance costs.