摘要
隧道衬砌中离散实体目标的检测精度和时效性直接关乎隧道的运营安全,采用图像视觉技术进行图像自动解译可极大提升检测效率和结果的准确性,因此基于离散实体目标的雷达图像数据构建自定义雷达数据集合,并提出一套改进的Faster R-CNN算法对隧道衬砌中的离散实体目标进行自动检测。该算法首先对现有Faster R-CNN网络的特征提取模块进行改进,提出一套全新的轻量化特征提取网络ResNet_FMBConv对雷达图像特征进行深度挖掘;基于ResNet_FMBConv网络改进现有特征金字塔(FPN)结构,实现对多尺寸下目标的精准辨识。其次,基于实测和仿真的雷达图像数据构建离散实体目标的自定义雷达数据集合,通过几何变换方法对雷达图像进行数据增强后用于算法验证。结果表明,改进算法在IOU=0.50∶0.95情况下的检测精确率、召回率、F 1分数和FPS分别为45.1%、54.0%、49.1%和21.65 fps。在保证召回率基本持平的情况下,同比YOLOv3_spp、SSD、Retinanet和Faster R-CNN等目标检测算法的精确率和F 1分数分别提升2%~9%和1%~6%。同时,试验结果表明改进后的特征提取网络ResNet_FMBConv也优于现有Resnet-50、VGG16、Efficientnet_b0和Mobilenetv3等目标分类网络。
As detection accuracy and timeliness of discrete entity objects in tunnel lining are directly related to the operational safety of the tunnel,the automatic image interpretation using image vision technology can greatly improve the efficiency and accuracy of detection results.Therefore,by constructing a custom GPR dataset based on the radar image data of discrete entity objects,this paper proposed an improved Faster R-CNN algorithm to automatically detect discrete entity objects in tunnel lining.Firstly,this algorithm improved the feature extraction module of the existing Faster R-CNN network,and a new lightweight feature extraction network ResNet_FMBConv was proposed to mine the radar image features in depth.Based on the ResNet_FMBConv network,the existing feature pyramid network(FPN)structure was improved to achieve accurate identification of objects under multiple sizes.Secondly,based on the measured and simulated radar image data,by constructing a custom GPR dataset of discrete entity objects,this paper used the geometric transformation method to enhance the radar image data for the algorithm verification.The results show that the detection accuracy,recall rate,balanced F 1-score and FPS of the improved algorithm are 45.1%,54.0%,49.1%and 21.65 fps respectively when IOU=0.50∶0.95.Under the condition that the recall rate remains basically unchanged,the precision rate and balanced F 1-score of target detection algorithms such as YOLOv3_spp,SSD,Retinanet and Faster R-CNN are improved by 2%~9%and 1%~6%,respectively.Meanwhile,the experimental results show that the improved feature extraction network ResNet_FMBConv is also superior to the existing target classification networks such as Resnet-50,VGG16,Efficientnet_b0 and Mobilenetv3.
作者
崔广炎
王艳辉
徐杰
丁冠军
秦湘怡
任秋阳
CUI Guangyan;WANG Yanhui;XU Jie;DING Guanjun;QIN Xiangyi;REN Qiuyang(State Key Laboratory of Advanced Rail Autonomous Operation,Beijing Jiaotong University,Beijing 100044,China;School of Traffic and Transportation,Beijing Jiaotong University,Beijing 100044,China;Beijing Engineering Research Center of Urban Traffic Information Intelligent Sensing and Service Technologies,Beijing Jiaotong University,Beijing 100044,China;Research and Development Center of Transport Industry of Technologies and Equipment of Urban Operation Safety Management,Beijing Jiaotong University,Beijing 100044,China)
出处
《铁道学报》
EI
CAS
CSCD
北大核心
2024年第2期171-180,共10页
Journal of the China Railway Society
基金
自然科学横向项目(I23L00060)。
