多维图像融合驱动的复杂场景路表破损识别  

Pavement surface distress detection in complex scenarios driven by multi-dimensional image fusion

作　　者：管进超 丁玲 杨旭 刘鹏飞[5] 汪海年 GUAN Jin-chao;DING Ling;YANG Xu;LIU Peng-fei;WANG Hai-nian(School of Highway,Chang'an University,Xi'an 710064,Shaanxi,China;Zhejiang Expressway Co.,Ltd.,Hangzhou 310020,Zhejiang,China;School of Transportation Engineering,Chang'an University,Xi'an 710064,Shaanxi,China;School of Future Transportation,Chang'an University,Xi'an 710064,Shaanxi,China;Institute of Highway Engineering,RWTH Aachen University,Aachen 52074,North Rhine-Westphalia,Germany)

机构地区：[1]长安大学公路学院,陕西西安710064 [2]浙江沪杭甬高速公路股份有限公司,浙江杭州310020 [3]长安大学运输工程学院,陕西西安710064 [4]长安大学未来交通学院,陕西西安710064 [5]德国亚琛工业大学道路工程研究所,北威州亚琛52074

出　　处：《交通运输工程学报》2024年第3期154-170,共17页Journal of Traffic and Transportation Engineering

基　　金：国家重点研发计划(2021YFB2601000);国家自然科学基金项目(52078049)。

摘　　要：为提升复杂场景中路表裂缝与坑槽的识别精度和鲁棒性,考虑实际检测场景中路表破损形态的不规则性和环境噪声干扰,提出了一种面向多维图像的路表破损自动分割模型与特征融合优化方法;基于多目立体视觉重构的路表高精度点云模型,通过同源点云栅格化生成二、三维图像,建立了复杂场景路表破损图像数据集;结合深度可分离卷积和多层位特征叠加,构造了轻量化编码-解码网络PDU-net,用于像素级裂缝与坑槽识别;在分割模型基础上,提出了像素运算和通道重组2种多维图像融合策略,以提升深度学习网络对浅细裂缝特征的提取效率。试验结果表明:PDU-net模型能够有效学习不同类型图像和破损特征,在不同数据集上的训练损失均能稳定收敛,其中三维图像训练周期小于二维图像;相较于现有卷积分割网络,PDU-net模型在复杂场景下的路表破损分割精度和效率更高,三维裂缝与坑槽图像分割的调和均值分别为81.00%和95.85%,平均正向推理时间约为现有模型的30%;多维融合图像可以提升复杂裂缝分割的精度和鲁棒性,在最优色彩-深度比为0.2时,裂缝分割的调和均值可提升至83.31%。综上所述,所提出的方法可在复杂场景中有效抑制环境噪声并强化病害特征。To improve the accuracy and robustness of crack and pothole detection of pavement surface in complex scenarios,the morphological irregularity of pavement surface distresses and the influence of environmental noises in practical detection scenarios were considered,and an automatic pavement surface distress segmentation model and feature fusion optimization method for multi-dimensional images were proposed.Based on high-precision pavement surface point cloud models reconstructed by multi-view stereo vision,2D and 3D images were generated by the rasterization of homologous point clouds.The pavement surface distress image dataset in complex scenarios was established.A lightweight encoding-decoding network,namely PDU-net,integrating depthwise separable convolution and multi-layer feature combination,was developed for pixel-level crack and pothole detection.Based on the segmentation model,two multi-dimensional image fusion strategies,including pixel operation and channel recombination,were proposed to improve the extraction efficiency of deep learning networks in shallow and fine crack features.Experimental results show that the PDU-net model can effectively learn features from different types of images and distresses.The training loss of the PDU-net on different datasets can converge stably,with the training cycles of 3D images shorter than that of 2D images.Compared with existing convolutional segmentation networks,the PDU-net model achieves higher accuracy and efficiency for pavement surface distress segmentation in complex scenarios.The harmonic means of 3D crack and pothole image segmentation are 81.00%and 95.85%,respectively.The average forward inference time of the PDU-net is about 30%of the existing models.The segmentation accuracy and robustness of complex cracks can be improved by multi-dimensional fusion images.When the optimal color-depth ratio is 0.2,the harmonic mean of the crack segmentation increases to 83.31%.In conclusion,the proposed method can effectively suppress environmental noises and strengthen

关 键 词：道路工程 路表破损识别 多维图像分割 复杂检测场景 轻量化编码-解码网络 图像特征融合 

分 类 号：U418.6[交通运输工程—道路与铁道工程]