摘要
针对概率模型方法不适用于大场景点云配准的问题,本文提出了一种点云局部主曲率特征约束的匹配方法,通过匹配下采样特征四元数,计算对应特征在原始点云的最近邻质心,对该点集使用连贯点漂移算法。斯坦福(Stanford)数据集的实验表明,本文算法相比现有主流方法,在小规模点云配准上有较快的速度和较高的精度,且对点云初始姿态要求不高,初始位姿差异较小时,在Bunny数据取得了6.073×10^(-3) mm的均方根误差;初始位姿较差时,迭代最近邻点算法(ICP)无效,本文算法取得了3.743×10^(-1) mm的均方根误差,在Dragon数据的均方根误差为1.639 mm。武汉大学地面站扫描点云配准基准(WHU-TLS)数据集的实验表明,本文算法可以自动为ICP算法提供良好的初值,能够推广至大场景点云配准应用。
The probabilistic model method is not suitable for point cloud registration in large-scale scenarios.To address this issue,this paper proposed a matching method based on local principal curvature feature constraints of point clouds.By matching the feature quaternions with point cloud down-sampling,the paper calculated the nearest neighbor centroids of the corresponding feature in the original point cloud and used a coherent point drift algorithm for the point sets.The experiments on the Stanford dataset show that the proposed method has faster speed and higher accuracy in small-scale point cloud registration compared to existing methods,and it has lower requirements for the initial pose of point clouds.The proposed method obtains root-mean-square error(RMSE)of 6.073×10^(-3) mm on Bunny data when the initial pose difference is small.When the initial pose is poor,the iterative closest point(ICP)algorithm is ineffective,and the proposed method obtains RMSE of 3.743×10^(-1) mm and 1.639 mm on Dragon data.The experiments on the WHU-TLS dataset show that the proposed algorithm can provide fine initial values for the ICP algorithm automatically and can be used for point cloud registration in large-scale scenarios.
作者
张银屏
董明
毛力妹
ZHANG Yinping;DONG Ming;MAO Limei(Jiangxi Ganhe Surveying and Mapping Geographic Information Company Limited,Shangrao,Jiangxi 334000,China)
出处
《北京测绘》
2025年第1期33-39,共7页
Beijing Surveying and Mapping
关键词
点云配准
连贯点漂移
局部主曲率
点云下采样
point cloud registration
coherent point drift
local principal curvature
point cloud down-sampling
