基于多目视觉深度神经网络的路面宏观纹理三维重建方法  

A Three-dimensional Reconstruction Method of Pavement Macro-texture Using a Multi-view Deep Neural Network

作　　者：马涛[1] 童峥 张一鸣 张伟光[1] MA Tao;TONG Zheng;ZHANG Yi-ming;ZHANG Wei-guang(School of Transportation,Southeast University,Nanjing 211189,Jiangsu,China)

机构地区：[1]东南大学交通学院,江苏南京211189

出　　处：《中国公路学报》2023年第3期70-80,共11页China Journal of Highway and Transport

基　　金：国家重点研发计划项目(2020YFB1600102);国家自然科学基金优秀青年科学基金项目(51922030)。

摘　　要：为提升路面宏观纹理三维重建的精度与效率,实现路面抗滑与抗磨耗性能的精准评估,提出了一种基于多目视觉深度神经网络的路面宏观纹理三维重建方法。首先,采用多目相机采集沥青路面多视角图像;其次,针对各个视角沥青路面图像,使用深度卷积神经网络提取与宏观纹理相关的高维特征向量,采用特征映射单元将特征向量映射为三维矩阵,使用多个反卷积层将三维矩阵转化为三维体素模型;最后,多目视觉组合模块采用贝叶斯规则融合不同视角的三维体素模型,该模型即为路面宏观纹理三维重建模型,可用于评估路面抗滑与抗磨耗性能。使用多目相机采集16条沥青路面的多目视觉图像,并用三维扫描仪采集三维点云数据,构建数据集验证该方法的准确性与稳定性。试验结果表明:多目视觉深度神经网络能准确地重建路面宏观纹理,50 DPI(Dots Per Inch)与70 DPI分辨下三维重建结果与三维点云数据的体积交并比分别为0.858、0.769,且交并比不受路面材料与背景噪音影响,具有高准确性与高稳定性;该方法的准确性与稳定性优于MVF-CNN、3D-FHNET、Stereo-vision等已有先进三维重建方法。路面宏观纹理三维重建结果可用于评估路面抗磨耗性能与抗滑性能,实测误差分别为6.82%、7.28%,精度满足路面性能检测需求。此外,基于路面三维宏观纹理的平均构造深度与动态抗滑系数的测试速度为60 km·h^(-1),具有高效性。路面宏观纹理三维模型未来可用于构建公路数字孪生体。To improve the accuracy and efficiency of the three-dimensional(3D)reconstruction of pavement macro-texture and to achieve high accurate evaluation of pavement anti-skid and anti-abrasive performances,a multi-view deep neural network has been proposed to perform pavement macro-texture reconstruction.First,a multi-view camera was used to collect pavement images with different perspectives.Then,a deep convolutional neural network was used to exact high-dimension features from each image,and the features were then mapped into a 3D matrix by a feature mapping unit.The 3D matrix was converted into a 3D voxel model by several deconvolution layers.Finally,the 3D voxel models in different views were combined by Bayesian rule,which was the 3D reconstruction result of pavement macro-texture,which was used to evaluate the anti-skid and anti-abrasive performances.The multi-view images and the 3D point cloud data were collected from 16 asphalt pavements by a multi-view camera and a 3D scanner,respectively.The two types of pavement data were combined to build a dataset to demonstrate the accuracy and stability of the proposed method.The test results showed that the multi-view deep neural network reconstructed the pavement macro-texture with the values of intersection over union(IoU)of 0.858 and 0.769 under 50 DPI and 70 DPI resolutions.The IoU results were not affected by the road surface material and background noise.The accuracy and stability of the proposed method were better than that of MVF-CNN,3D-FHNET and stereo-vision methods.The 3D reconstruction results were used to evaluate mean texture depth(MTD)and dynamic friction coefficient(DFC)with the measured error is 6.82%and 7.28%,respectively,which meet the requirements of pavement performance detection.In addition,the test speed of MTD and DFC was 60 km·h^(-1),which had high efficiency.The 3D model of pavement macro-texture can be used to build the highway digital twin in the future.

关 键 词：道路工程 路面宏观纹理 深度神经网络 多目视觉 目标三维重建 

分 类 号：U416.2[交通运输工程—道路与铁道工程]