基于线特征约束ICP算法的机载激光雷达点云与影像配准研究  

Research on Airborne LiDAR Point Cloud and Image Registration Based on ICP Algorithm with Line Feature Constraints

作　　者：赵思亮 ZHAO Siliang(China Railway Design Corporation,Tianjin 300251,China)

机构地区：[1]中国铁路设计集团有限公司,天津300251

出　　处：《铁道勘察》2024年第5期84-89,95,共7页Railway Investigation and Surveying

基　　金：中国铁路设计集团有限公司科技研发计划重点课题(2023A0253820)。

摘　　要：传统ICP算法存在计算效率低、易陷入局部最优等缺点,故提出一种附有线特征约束的ICP算法。该方法采用由粗到精的配准策略,首先在点云和影像数据生成的DSM图像中使用SIFT算法进行特征点配对,经RSANC算法消除误匹配后,作为数据配准的初始参数;然后分别在原始点云数据和影像DSM数据中提取建筑物边界线作为线特征,并将线特征进行简化、拟合、正交化;而后用端点表示特征线,将同名特征线的端点加入到ICP算法中作为线特征约束,完成点云数据和影像密集点云数据的配准。相较于传统ICP算法,新方法不需要对全部的点云数据进行运算,只需抽样少量点(3%左右)参与配准计算,即可完成点云与影像3D-3D配准,算法效率大幅提高,同时在线特征约束下,可避免传统算法容易陷入局部最优的缺点。实验表明,改进算法将配准结果中平均点间距由0.462 m提高到0.342 m,精度提升显著。The traditional ICP algorithm has drawbacks such as low computational efficiency and susceptibility to local optima.This paper added line feature constraints to the traditional ICP algorithm and proposed an ICP algorithm with line feature constraints.The method adopted a coarse to fine registration strategy.Firstly,the SIFT algorithm was used to pair feature points in the DSM images generated from point cloud and image data.After eliminating mismatches with the RSANC algorithm,it served as the initial parameters for data registration.Then building boundary lines were extracted as line features from the original point cloud data and image DSM data,and the line features were simplified,fitted,or orthogonalized.The endpoints were used to represent the feature lines,and the endpoints of the same named feature lines were added to the ICP algorithm as line feature constraints to complete the registration of point cloud data and image dense point cloud data.Compared with the classic ICP algorithm,the new method proposed in this article fully extracts point cloud and image features without the need for operations on all point cloud data.Only a small number of points(about 3%)need to be sampled to participate in registration calculation,which can complete the registration of point cloud and image 3D-3D.The computational efficiency is greatly improved,and under the constraint of online features,it avoids the disadvantage of traditional algorithms easily falling into local optima.Through data experiments,the improved algorithm significantly increased the average point spacing of the painful ICP algorithm registration results from 0.462 m to 0.342 m,resulting in a significant improvement in accuracy.

关 键 词：机载激光雷达 特征提取 配准 ICP算法 点云 

分 类 号：U212[交通运输工程—道路与铁道工程] P221[天文地球—大地测量学与测量工程]