Abstract: Fires in tunnels have the problem of difficult detection and rescue. Thus real-time fire monitoring is crucial for prompt discovery of fires. Traditional video image-based fire detection methods rely on a single image and cannot extract temporal and spatial information from multiple images. The detection precision is limited and not effective in detecting tunnel fires. As a result, a tunnel fire frame difference network is proposed. The structure of the frame difference network is mainly a 3D convolutional kernel. Temporal contextual information about the fire in the video can be extracted. The network connects to the YOLOv5 backbone network to form tunnel fire detection network, which can detect a single image as well as two images. This takes full advantage of the motion information of the videos. The network achieves fast convergence by optimizing the bounding box loss using the CIoU function and combining classification loss and confidence loss. Experimental results show that the network achieves an average precision of 91.03% on the tunnel fire dataset, with a detection speed of 63.7 frames per second. The model has strong robustness. The network and analysis strategies are combined to form a tunnel fire detection system. It has a false detection rate and a missed detection rate of 2.52% and 2.03% in the tunnel scene, respectively. The results can satisfy the accurate and real-time requirements of tunnel fire detection.

Key words: tunnel, fire detection, frame difference network, YOLOv5

摘要： 隧道发生火灾存在着检测难、救援难的问题，实时的火灾监测对于及时发现火情是至关重要的。传统基于视频图像的火灾检测方法，检测依赖单幅图像，无法提取多幅图像的时空信息，检测精度低，不能有效检测隧道火灾。因此，提出了隧道火灾帧差网络。帧差网络使用3D卷积核构建网络结构，提取视频中火灾的时间上下文信息；将帧差网络衔接至YOLOv5主干网络形成隧道火灾帧差检测网络，可以检测单幅图像及两幅图像，从而充分利用视频动态信息；使用CIoU函数优化网络的边界框损失，并融合分类损失与置信度损失，使网络能够快速收敛。实验结果表明，该网络在隧道火灾数据集上的平均精度高达91.03%，检测速度达到了63.7帧/s，具有较强的鲁棒性。通过选取最优分析策略设计隧道火灾检测应用方法，该方法在隧道场景中的漏检率和误检率分别为2.52%和2.03%，可以满足隧道火灾检测的准确性和实时性需求。

关键词: 隧道, 火灾检测, 帧差网络, YOLOv5