基于改进YOLOv5s的公路隧道火灾烟雾智能检测算法  

Intelligent Detection Algorithm for Fire Smoke in Highway Tunnel Based on Improved YOLOv5s

作　　者：邓实强 丁浩[2] 蒋树屏 杨孟[2] 刘帅 陈建忠[2] 李文锋[2] DENG Shi-qiang;DING Hao;JIANG Shu-ping;YANG Meng;LIU Shuai;CHEN Jian-zhong;LI Wen-feng(School of Civil Engineering,Chongqing Jiaotong University,Chongqing 400074,China;Research Institute of Tunnel and Underground Engineering,China Merchants Chongqing Communications Technology Research&Design Institute Co.Ltd.,Chongqing 400067,China)

机构地区：[1]重庆交通大学土木工程学院,重庆400074 [2]招商局重庆交通科研设计院有限公司隧道与地下工程研究院,重庆400067

出　　处：《中国公路学报》2024年第11期194-209,共16页China Journal of Highway and Transport

基　　金：国家重点研发计划项目(2021YFC3002000);重庆市自然科学基金项目(CSTB2022NSCQ-MSX1049)。

摘　　要：隧道火灾快速感知可为隧道运营安全提供重要保障,并为隧道应急处置提供关键性决策信息。然而,现有的视频图像火灾烟雾检测方法在公路隧道复杂环境下存在准确性和时效性问题,并且缺乏基础视频图像数据。为此,通过开展实体公路隧道火灾试验,创建高清视频图像数据集,以真实隧道场景下模拟的火灾烟雾视频图像为研究对象,提出一种基于改进YOLOv5s的智能检测算法。在模型中,使用增强后的Mosaic方法对训练数据进行增强处理;引入Transformer Encoder模块增强网络全局特征提取的能力,改善较小烟雾目标特征提取较为困难的状况,以提升网络性能;利用最新的轻量级卷积方法GSConv替换掉部分卷积模块Conv,减少网络参数的同时保持网络性能,达到压缩网络目的;添加轻量级高效通道注意力模块ECA,通过局部跨通道交互策略缓解监控摄像机距目标较远与火灾初期烟雾漏检问题,在少参数量增加的情况下进一步提升网络性能;采用CIoU损失函数与SiLU激活函数的组合使网络更快得到收敛。为验证所提算法的有效性,选用YOLOv3、YOLOv3-efficientnet、YOLOv5s、YOLOX、YOLOv7、YOLOv7-tiny、SSD七种目标检测算法进行对比分析。结果表明:在自建的公路隧道火灾烟雾数据集上,所提算法的检测精确度达到97.27%,mAP@0.5为97.85%。尽管在原网络基础上提升的幅度仅为1.83%和1.13%,但相较于其他7种对比算法,所提算法对远距离和火灾初期的小尺寸烟雾目标有更好的检测效果,明显改善了漏检情况。此外,算法的检测速度为86.2帧·s~(-1),能够满足隧道火灾检测的时效性要求,同时利用重庆真武山隧道火灾视频验证了算法的可靠性,研究结果可为实现隧道复杂环境下的火灾快速感知提供技术支持。The rapid detection of tunnel fires provides an important guarantee of tunnel operation safety and critical decision-making information for tunnel emergency responses.However,existing fire smoke detection methods using video image have issues regarding accuracy and timeliness in complex highway tunnel environments,and lack basic video image data.Therefore,in this study,physical highway tunnel fire experiments were conducted to obtain high-definition video image datasets considering smoke video images simulated under real tunnel scenarios as the research object.An intelligent detection algorithm based on improved YOLOv5s was realized.The enhanced mosaic method is incorporated for training data augmentation,while the Transformer Encoder module is introduced to enhance the network's global feature extraction capabilities as well as improve the problem of difficult feature extraction for small smoke targets,thereby improving the network performance.The latest lightweight convolution method,GSConv,was further utilized to replace some of the Conv modules,reducing the network parameters while maintaining network performance to achieve network compression.Furthermore,a lightweight efficient channel attention(ECA)module is added to alleviate issues such as missed detections of smoke targets at long distances and small targets using a local cross-channel interaction strategy,further improving network performance without increasing the number of parameters.A combination of the CIoU loss function and SiLU activation function are considered to allow the network to converge more quickly.To verify the effectiveness of the proposed algorithm,seven target-detection models were selected for comparison and analysis:YOLOv3,YOLOv3-efficientnet,YOLOv5s,YOLOX,YOLOv7,YOLOv7-tiny,and SSD.The results indicate that regarding the self-built highway tunnel fire smoke dataset,the proposed algorithm achieves the detection accuracy of 97.27%and mAP@0.5 of 97.85%.Although the improvements over the original network are only 1.83%and 1.13%,respect

关 键 词：隧道工程 烟雾检测 火灾试验 目标检测 YOLOv5s 

分 类 号：U458.1[建筑科学—桥梁与隧道工程]