基于深度学习和虚拟模型的路面全域伤损状态自动化感知  

Automated Pavement-wide Injury State Sensing Based on Deep Learning and Virtual Models

作　　者：邱实 陈斌 胡文博[1,3] 王卫东 伍定泽[1,3] 张晨雷 汪雯娟 高洪波 王劲 QIU Shil;CHEN Bin;HU Wen-bol;WANG Wei-dong;WU Ding-ze;ZHANG Chen-leil;WANG Wen-juan;GAO Hong-bo;WANG Jin(School of Civil Engineering,Central South University,Changsha 410075,Hunan,China;Transportation Department of Guangxi Zhuang Autonomous Region,Nanning 530000,Guangxi,China;MOE Key Laboratory of Engineering Structures of Heavy-haul Railway,Central South University,Changsha 410075,Hunan,China;School of Business Administration,Capital University of Economics and Business,Beijing 100026,China;Guizhou Guijin Expressway Co.Ltd.,Guiyang 55008l,Guizhou,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075 [2]广西壮族自治区交通运输厅,广西南宁530000 [3]中南大学重载铁路工程结构教育部重点实验室,湖南长沙410075 [4]首都经济贸易大学工商管理学院,北京100026 [5]贵州贵金高速公路有限公司,贵州贵阳550081

出　　处：《中国公路学报》2023年第3期61-69,共9页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52178442);贵州贵金高速公路有限公司开发项目(GJGS-2020-ZX-076)。

摘　　要：路面全域伤损的有效感知是全面、系统地实施维护决策的关键依据。提出一种基于深度学习和虚拟模型的路面全域伤损状态自动化感知方法。该方法首先基于无人机实测数据建模生成路面虚拟实体;然后使用深度语义分割网络从实测数据中精细化地检测路面伤损;最后,将输出的伤损特征检测结果与虚拟实体数据进行匹配和UV映射,获取各个伤损在虚拟空间中的定位信息并逐一部署,得到面向路面全域的伤损状态感知模型。结果表明:在现场实测统计长为236 m,宽为20 m的实际路面区域试验研究中,充分训练的U-Net网络的平均交并比(MIoU)达到0.86,显示出对无人机采集到的路面伤损区域极佳的分割精度。建立的路面伤损状态感知模型有效感知实际存在的伤损91处,与传统的二维检测结果相比,能够更加系统地对路面全域伤损进行全局表征,便于高效地推断伤损的特征和位置信息,实现精准、动态的路面服役状态评估。Effective measuring of full pavement performance is the key basis for comprehensive and systematic implementation of maintenance decisions.In this paper,we propose an automated method for pavement full-area performance measuring based on deep learning and virtual model.The method first generates pavement virtual entities based on UAV real measurement data modeling;then uses a deep semantic segmentation network to finely detect pavement performance from real measurement data;finally,the output performance feature detection results are matched and UV mapped with virtual entity data to obtain the positioning information of each performance in the virtual space and deploy them one by one to obtain damage state sensing model for the full area of the pavement.The results show that the average cross-merge ratio(MIoU)of the fully trained U-Net network reaches 0.86,showing an excellent segmentation accuracy of the pavement performance areas collected by the UAV.The pavement damage state sensing model established in this paper effectively senses 91 actual existing injuries in a pavement area of 236 m in length and 20 m in width,which enables a more systematic global characterization of pavement-wide injuries compared with the traditional two-dimensional inspection results,facilitating efficient inference of injury characteristics and location information,and realizing accurate and dynamic pavement service state assessment.

关 键 词：道路工程 路面伤损检测 深度学习 状态感知模型 

分 类 号：U418.6[交通运输工程—道路与铁道工程]