基于流形聚类的欠采样非均匀密度三维点云配准

Undersampled non-uniform density multi-station 3D point cloud alignment method based on manifold clustering

三维点云配准中各个视图的激光扫描点云数据彼此之间会存在重叠的部分。针对由于重叠区域不同而造成的点云密度不均匀直接影响多站点云配准精度的问题,本文提出一种基于流形聚类的欠采样非均匀密度多站三维点云配准方法。首先,将测地距离作为相似性度量,对不平衡点云数据进行聚类划分实现点云数据精简;然后,采用K邻域搜索方法计算每个点半径范围内点的个数,划分点云分为密度区域;再对密度较大的区域进行聚类并对每个聚类进行曲面拟合,计算曲面上所有点的曲率;再提取曲率较大的点,使得密度较大的区域与密度较小的区域中点云的个数基本保持平衡,得到密度较为平衡的点云数据。最后,将流形聚类欠采样后的点云使用K均值(K-means)进行聚类并更新聚类中心和刚性变换矩阵,实现非均匀密度多站点云配准。与随机采样法和均匀采样法相比,本文方法的倒角距离较小,并且保留了点云的局部特征信息。在斯坦福大学公共数据集中的Bunny数据集上的实验表明,所提方法在保证配准精度的前提下使配准的效率提高了60%以上。

An under-sampled non-uniform density multi-station 3D point cloud alignment method based on manifold clustering is proposed,to address the problem that the point cloud data from each view overlap with each other,and the uneven point cloud density caused by different overlapping areas directly affect the multi-station cloud alignment accuracy.First,the geodesic distance is used as a similarity measure to cluster the unbalanced point cloud data to achieve a streamlined point cloud data.Then,the K nearest neighbour(KNN)method is used to calculate the number of points within the radius of each point,and the point cloud is divided into denser and less dense point clouds.Next,the denser regions are clustered and the surfaces are fitted to each cluster,and the curvatures of all points on the surfaces are calculated.The points with greater curvature are extracted,so that the denser regions and the points with greater curvature are extracted so that the number of point clouds in the denser regions and the less dense regions are balanced,resulting in more balanced point cloud data.Finally,the point clouds are undersampled using manifold clustering and clustered using K-means clustering,which updates the clustering centres and the rigid transformation matrix to achieve non-uniform density multi-station cloud alignment.Compared with the random sampling method and the uniform sampling method,the proposed method has a smaller chamfer distance and preserves the local feature information of the point cloud.The experiments on the Bunny dataset in the Stanford University public dataset indicate that the proposed method improves the alignment efficiency by more than 60%while ensuring the accuracy of the alignment.