基于特征点提取的RANSAC-ICP三维点云配准方法  

RANSAC-ICP 3D point cloud registration method based on feature point extraction

作　　者：李岩 张子毅 王建柱 LI Yan;ZHANG Ziyi;WANG Jianzhu(Shandong Expressway Construction Management Group Co.,Ltd.,Jinan 250001,Shandong,China;School of Oilu Transportation,Shandong University,Jinan 250002,Shandong,China)

机构地区：[1]山东高速建设管理集团有限公司,山东济南250001 [2]山东大学齐鲁交通学院,山东济南250002

出　　处：《山东大学学报（工学版）》2024年第5期144-154,共11页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金资助项目(52002224);山东省重点研发计划重大科技创新工程(2020CXGC10118);国家重点研发计划资助项目(2022YFB2602102)。

摘　　要：针对现有点云配准算法中易出现的误匹配、迭代时间长、精度低等问题,提出一种基于特征点提取的随机采样一致性与迭代最近点的三维点云配准方法,设计体素滤波降采样、关键特征点提取和几何特征描述、改进的随机采样一致性和点到面的迭代最近点算法框架。在降采样的基础上,提取点云关键几何特征点,并进行关键点邻域描述。采用四点对的随机采样一致性算法和点到面的迭代最近点算法分别进行点云粗配准与精细配准;采用K维树方法加速迭代,奇异值分解求解最优变换矩阵,最终实现三维点云配准。利用激光雷达及配套设备,模拟无人车辆采集点云数据,并选取不同时间间隔的点云图像,引入均方根误差和运算时间指标,验证算法的性能。试验结果表明:在粗配准阶段,配准速度相较于采样一致性算法和四点一致性算法,配准速度平均提高78.44%和61.02%,在处理100帧以下的数据时,配准误差在10 cm范围内;在精配准阶段,配准误差较粗配准、正态分布变换算法、传统的迭代最近点算法分别降低5.11、4.94和0.53 cm,配准时间较传统的迭代最近点算法平均提高33.06%。Aiming at solving the problems of mismatch,long iteration time and low accuracy in data registration,this paper proposed a random sample consensus-iterative closest point registration method based on feature point registration.The algorithm framework consisted of voxel filtering down-sampling,key feature point extraction and geometric feature description,improved random sample consensus and point-to-plane iterative closest point iteration.Based on downsampling,the key geometric feature points of the point cloud were extracted and the key point neighborhood description was performed.We used the random sample consensus algorithm for four-point pairs and the point-to-plane iterative closest point algorithm to realize the point cloud coarse alignment and fine alignment.LiDAR and supporting equipment were used to simulate unmanned vehicles to collect point cloud data.According to the experimental data,the point cloud images of different time intervals were selected,and the root mean square error and the operation time index were introduced to verify the performance of the algorithm.In terms of coarse registration,the registration speed was improved by 78.44%and 61.02%compared with sample consensus initial alignment and 4-Points Congruent Sets algorithms,and the registration error below 100 frames was within 10 cm.In terms of fine registration,the registration error was 5.11 cm,4.94 cm and 0.53 cm lower than that of coarse registration,normal distributions transform algorithm and traditional iterative closest point algorithm respectively,and the registration time was 33.06%higher than that of traditional iterative closest point algorithm.

关 键 词：激光SLAM 点云配准 关键点提取 随机采样一致性 迭代最近点 

分 类 号：TN958.98[电子电信—信号与信息处理]