Tunnel Fire Detection Based on Improved RT-DETR

With the growing frequency of tunnel fire incidents globally, this paper designs an improved RT-DETR model to achieve real-time and precise fire detection. First, within the backbone of RT-DETR, the FasterEMA Block is employed to replace the original BasicBlock, thereby enhancing feature extraction capabilities. Second, in the feature pyramid, a MixerFFN Layer is designed to replace the AIFI module to strengthen feature fusion performance. Finally, the LSKA attention mechanism is utilized in the path aggregation network to optimize the receptive field adjustment of feature maps and to improve detection accuracy for small targets. Experimental results show that the improved RT-DETR model achieves a mean Average Precision (mAP) of 0.917, an increase of 2.68% over the baseline model, while simultaneously reducing the GFLOPs by 7.21%. This accomplishes a dual enhancement in both detection accuracy and real-time performance.