嵌入Dense Net的YOLOv4多尺度隧道火灾检测算法

Yolov4 Multi-Scale Tunnel Fire Detection Algorithm Embedded In DenseNet

为解决现有火灾检测算法在公路隧道的应用较少以及算法检测的准确性和实时性不能满足要求,提出一种改进的YOLOv4隧道火灾检测算法。由于公开的隧道火焰数据集较少,通过模拟火灾形成初期的火焰和网上搜集构建了一个包含3000张隧道火焰图片的数据集,在数据集中添加了500张隧道车灯和500张隧道照明灯用于模型训练,提高泛化性。然后在YOLOv4的特征提取层的CSP结构中嵌入DenseNet,融合多尺度的浅层特征信息,增强特征复用,降低梯度消失的问题和抑制过拟合。再在特征网络检测层中增加了104×104特征检测尺寸,增强隧道火灾形成初期小目标火焰的检测精度。实验结果表明:上述算法检测精度为90.4%,在原YOLOv4网络上提高了5.2%,其检测速度可达0.16s,满足隧道实时检测的要求,并且对小尺寸火焰也有很好的检测效果,对实现隧道火灾初期预警具有重要意义。

In order to solve the problem that the existing fire detection algorithms are less used in highway tunnels and the accuracy and real-time performance of the algorithm detection cannot meet the requirements,an improved YOLOv4 tunnel fire detection algorithm is proposed.Due to the lack of public tunnel flame data sets,a data set containing 3000 tunnel flame pictures was constructed by simulating the flames at the initial stage of the fire formation and online collection,and 500 tunnel lights and 500 tunnel lights were added to the data set.For model training,improve generalization.Then the DenseNet network was embedded in the CSP structure of the feature extraction layer of YOLOv4,and the multi-scale shallow feature information was integrated to enhance feature reuse,reduce the problem of gradient disappearance and suppress over-fitting.Then,104x104 feature detection size was added to the feature network detection layer to enhance the detection accuracy of small target flames in the early stage of tunnel fire formation.The experimental results show that the detection accuracy of the algorithm is 90.4%,which is improved by 5.2%on the original YOLOv4 network,and its detection speed can reach 0.16s,which meets the requirements of real-time detection of tunnels,and has a good detection effect for small flames.Early warning of tunnel fire is of great significance.