基于E-HRNet的路面破损区域识别方法  

Road surface damage area identification method based on E-HRNet

作　　者：张大伟 田抑阳 徐培娟 钟琛 ZHANG Dawei;TIAN Yiyang;XU Peijuan;ZHONG Chen(School of Automobile,Chang’an University,Xi’an 710018,China;School of Transportation Engineering,Chang’an University,Xi’an 710018,China;Chongqing Rail Transit(Group)Co.,Ltd.,Chongqing 401120,China)

机构地区：[1]长安大学汽车学院,西安710018 [2]长安大学运输工程学院,西安710018 [3]重庆市轨道交通(集团)有限公司,重庆401120

出　　处：《北京交通大学学报》2023年第4期110-119,共10页JOURNAL OF BEIJING JIAOTONG UNIVERSITY

基　　金：陕西省自然科学基金(2023-JC-YB-496);中国博士后科学基金(2021M693752);轨道交通基础设施性能监测与保障国家重点实验室开放课题(HJGZ2022115);中央高校基本科研业务费专项资金(300102342104)。

摘　　要：针对现有路面破损区域识别方法识别效率低、泛化性能差等问题,提出了一种基于深度高分辨率网络HRNet的路面坑槽和裂缝识别方法.采用车载单目相机实地采集路面坑槽和裂缝图像,并对图像进行预处理和标注,生成路面破损数据集.在原HRNet基础上,对其网络特征提取层的4个不同分辨率表征分别融合改进的卷积注意力模块,形成E-HRNet网络模型.为了提高EHRNet模型的推理速度,对其各步骤中不同分辨率分支的残差层数进行了优化,并采用联合损失函数对该模型进行监督训练.试验结果表明:E-HRNet网络模型对路面坑槽和裂缝区域分割的平均像素精度和平均交并比分别达到了94.53%和88.31%,与原HRNet网络模型相比,平均像素准确率增加了6.53%,平均交并比提升了5.38%,平均类别准确率提高了1.39%;模型检测帧率提高了30.3%,而模型体积则减少了42.6%,可满足模型轻量化和实时检测的需求;与DDRNet、Deep⁃labV3+等同类模型相比,E-HRNet网络模型对坑槽和裂缝区域的分割精度更高,有效地避免了漏检、误检以及边界模糊等问题的出现,具有更好的实时性和泛化性.This paper proposes a method for identifying road potholes and cracks on road based on the depth high-resolution network HRNet,addressing the issues of low recognition efficiency and poor generalization performance in existing road damage identification methods.The method utilizes an onsite collection of road pothole and crack images using a vehicle-mounted monocular camera,followed by image preprocessing and annotation to generate a road damage dataset.Based on the original HRNet,the four different resolution representations in its network feature extraction layer are respec⁃tively fused with the improved convolution attention module to form the E-HRNet network model.To improve the reasoning speed of the E-HRNet model,the number of residual layers of different resolu⁃tion branches in each step is optimized,and the model is supervised and trained using a joint loss func⁃tion.Experimental results show that E-HRNet network model achieves an average pixel accuracy and average intersection over union acquired of 94.53%and 88.31%,respectively,for road pothole and crack segmentation.Compared with the original HRNet network model,E-HRNet model shows im⁃provements of 6.53%in average pixel accuracy,5.38%in average intersection over union,and 1.39%in average class pixel accuracy.Additionally,the E-HRNet model exhibits a 30.3%improve⁃ment in model detection frame rate,and a 42.6%reduction in model volume,meeting the require⁃ments of lightweight and real-time detection.Furthermore,compared with similar models such as DDRNet and DeeplabV3+,the E-HRNet network model achieves higher segmentation accuracy for potholes and cracks,effectively addressing issues of missing detection,false detection and blurred boundaries,while demonstrating better performance and generalization.

关 键 词：视觉感知 E-HRNet 路面破损识别 路面坑槽 裂缝 

分 类 号：U495[交通运输工程—交通运输规划与管理]