Abstract: Point cloud segmentation is the basis of point cloud recognition and modeling. In order to improve the accuracy and efficiency of point cloud segmentation, a super voxel based on semiautomatic method for region growing segmentation of three-dimensional point cloud is proposed. First of all, the original point cloud is divided into super voxels by using Octree according to the spatial location and normal vector of the point clouds. Secondly, a new uniform density point cloud is built by using all the central units of super voxels. In this case, the number of original point clouds decreases, which leads to a reduced processing time. Then, K-D tree is utilized as data structure to organize the resampled cloud points, and the points are merged into clusters constrained by local connectivities and smoothness. Finally, all the segmented point clouds are up-sampled back to the original point clouds. The proposed methodology is implemented on LiDAR point clouds from three?plant phenological periods. The quantitative assessment results demonstrate that the average fitting degree of the segmented point cloud through the proposed methodology and the manual segmentation reaches 93.38%, and the results also show high accuracy and high efficiency.

Key words: super voxels, normal vector, region growing, point cloud segmentation

摘要： 点云分割是点云识别与建模的基础。为提高点云分割准确率和效率，提出一种结合超体素和区域增长的自适应分割算法。根据三维点云的空间位置和法向量信息，利用八叉树对点云进行初始分割得到超体素。选取超体素的中心体素组成一个新的重采样后的密度均匀点云，降低原始点云数据处理量，从而减少运算时间。建立重采样后点云数据的K-D树索引，根据其局部特征得到点云簇。最后将聚类结果返回到原始点云空间。分别选取植物三个物候期的激光扫描点云，对该方法的有效性进行验证。实验结果表明，该方法分割后点云与手工分割平均拟合度达到93.38%，高于其他同类方法，且算法效率得到明显提升。

关键词: 超体素, 法向量, 区域增长, 点云分割