基于雷视融合的轨道三维点云重构研究  

Research on 3D point cloud reconstruction of railway tracks based on LiDAR-vision fusion

作　　者：何庆[1,2] 付彬 王启航[1,2] 曾楚琦 郝翔 王平 袁泉 HE Qing;FU Bin;WANG Qihang;ZENG Chuqi;HAO Xiang;WANG Ping;YUAN Quan(MOE Key Laboratory of High-speed Railway Engineering,Southwest Jiaotong University,Chengdu 610031,China;School of Civil Engineering,Southwest Jiaotong University,Chengdu 610031,China;Guangzhou Metro Design and Research Institute Co.,Ltd.,Guangzhou 510000,China)

机构地区：[1]西南交通大学高速铁路线路工程教育部重点实验室,成都610031 [2]西南交通大学土木工程学院,成都610031 [3]广州地铁设计研究院股份有限公司,广州510000

出　　处：《北京交通大学学报》2024年第5期69-77,共9页JOURNAL OF BEIJING JIAOTONG UNIVERSITY

基　　金：国家自然科学基金(U1934214,51878576);中铁第一勘察设计院集团有限公司科技开发项目(2021KY20ZD(ZNGT)-09PT);国家重点研发计划(2017YFB1201102)。

摘　　要：针对激光雷达无法采集真实色彩信息、图像三维重构点云精度低等单一传感器面临的问题,提出一种融合激光点云与图像对轨道进行三维重建的方法.首先,通过平滑和映射紧耦合的激光雷达惯性里程计(Lidar Inertial Odometry via Smoothing and Mapping,LIO-SAM)进行实时轨道激光点云建图;然后,利用尺度不变特征变换(Scale Invariant Feature Transform,SIFT)算法提取多幅图像上的特征点,并通过匹配相同特征点的方式计算多视角图像之间的几何关系,由运动结构恢复(Structure From Motion,SFM)和多视角密集匹配(Multi-View Stereo,MVS)算法寻找、聚簇和生成包含轨道纹理色彩信息的稠密图像点云;最后,将轨道板的平面特征和钢轨线性特征作为索引特征,采用迭代最近点(Iterative Closest Point,ICP)算法将图像点云与激光点云进行合并配准,并以激光点云空间位置信息为基准,融合图像点云纹理色彩信息得到精准且真实感强的轨道三维模型.研究结果表明:相较于传统配准算法,改进算法的形状参数和最近邻点分布指标分别提升83.4%和85.9%;对轨道点云进行目标识别时,融合点云的总体精度较原始点云提升7.7%,在平均精度和均值交并等指标上表现更优;通过轨道融合点云计算得到的轨距、高差与实测数据的对比误差在3 mm以内,证明了轨道三维点云重构方法的有效性.To address the limitations of single sensors,such as the inability of LiDAR to capture true color information and the low accuracy of image-based 3D reconstruction point clouds,this paper proposes a method for 3D track reconstruction by integrating LiDAR point cloud data and image data.First,real-time laser point cloud mapping of the track is achieved using LiDAR Inertial Odometry via Smoothing and Mapping(LIO-SAM),a tightly coupled radar-inertial odometry method.Next,the Scale Invariant Feature Transform(SIFT)algorithm is employed to extract feature points from multiple images,and the geometric relationships between multi-view images are determined by matching corresponding feature points.Structure From Motion(SFM)and Multi-View Stereo(MVS)algorithms are then applied to locate,cluster,and generate dense image-based point clouds enriched with texture and color information of the track.Finally,plane features of the track slabs and linear features of the rails are used as reference features,and the Iterative Closest Point(ICP)algorithm is employed to merge and register the image point cloud with the LiDAR point cloud.By using the spatial position information of the LiDAR point cloud as a baseline and fusing it with the texture and color information from the image point cloud,an accurate and realistic 3D track model is obtained.Experimental results demonstrate that,compared with traditional registration methods,the improved algorithm achieves an 83.4%and 85.9%enhancement in shape parameter accuracy and nearest-neighbor point distribution,respectively.When performing target recognition on the track point cloud,the fused point cloud improves overall accuracy by 7.7%over the original point cloud and performs better on metrics such as average precision and mean intersection-over-union.Furthermore,the comparison between the calculated track gauge and height difference from the fused point cloud and the measured data reveals an error margin within 3 mm,verifying the effectiveness of the proposed 3D track point cloud recon

关 键 词：铁路轨道建模 雷视融合 数据融合 轨道点云模型 

分 类 号：U216.3[交通运输工程—道路与铁道工程]