基于CBAS_Unet的无人机多视角影像钢轨线自动提取方法  

Automatic Extraction of Rail Lines from Multi-View Images of UAV Based on CBAS_Unet

作　　者：王广帅 WANG Guangshuai(China Railway Design Corporation,Tianjin 300251,China;Tianjin Key Laboratory of Rail Transit Navigation Positioning and Spatio-Temporal Big Data Technology,Tianjin 300251,China)

机构地区：[1]中国铁路设计集团有限公司测绘地理信息研究院,天津300251 [2]天津市轨道交通导航定位及时空大数据技术重点实验室,天津300251

出　　处：《中国铁道科学》2024年第3期26-37,共12页China Railway Science

基　　金：国家重点研发计划项目(2022YFC3005201);天津市科技计划项目(23ZGSSSS00010);中国国家铁路集团有限公司科技研究开发计划课题(L2023G014);中国铁路设计集团有限公司科技开发重点课题(2023A0240109)。

摘　　要：针对无人机影像场景复杂、视角差异大、异物遮挡多,因而难以实现影像钢轨线高精度自动提取的问题,提出基于CBAS_Unet的无人机多视角影像钢轨线自动提取方法。在传统U-net网络基础上增加并行的空洞空间金字塔模块(Atrous Spatial Pyramid Pooling,ASPP)及卷积注意力模块(Convolutional Block Attention Module,CBAM),以增强网络对不同尺度的邻域信息的获取能力,有效提升对钢轨的分割性能;再通过基于RANSAC最小二乘拟合的像素编组及邻域钢轨线串联,实现完整钢轨矢量线的高精度提取。结果表明:与Unet及Deeplab v3+2种经典模型相比,所提方法针对多视角无人机影像钢轨分割的交并比分别提升2.09%和1.98%,综合能力评价指标分别提升1.50和1.42;钢轨线提取完整度达到了90.7%,优于U-net模型的83.3%;钢轨线提取的误差平均值约0.58像素,中误差约0.77像素,实现了亚像素级的钢轨线提取。该方法能够满足无人机多视角影像中钢轨线提取的自动化、完整性以及高精度应用的需求。Addressing the challenges posed by the complexity of UAV image scenes,significant differences in view angles,and frequent occlusions by foreign objects,an automatic extraction method of UAV multi-view image rail lines based on CBAS_Unet is proposed.On the basis of the traditional U-net network,a parallel Atrous Spatial Pyramid Pooling(ASPP)and Convolutional Block Attention Module(CBAM)are added to enhance the network’s capability to capture contextual information across various scales,thereby significantly boosting the segmentation performance of rails.High-precision extraction of complete rail vector lines is then achieved by pixel grouping using RANSAC least squares fitting and chaining of neighboring rail lines.The experimental results show that,compared with the two classical models of U-net and Deeplab v3+,the proposed method achieves an increase in the intersection and merger ratio for rail segmentation of multi-view UAV images by 2.09%and 1.98%,and the score value is improved by 1.50 and 1.42,respectively.The completeness of rail line extraction reaches 90.7%,surpassing the 83.3%achieved by the U-net model.The average error in rail line extraction is approximately 0.58 pixels,with the median error of approximately 0.77 pixels,enabling sub-pixel-level rail line extraction.This method fulfills the requirements for automation,comprehensive and high-precision extraction of rail lines from UAV multi-view images.

关 键 词：无人机影像 钢轨线提取 空洞空间金字塔 注意力机制 U-net 

分 类 号：U216.3[交通运输工程—道路与铁道工程]