基于法向量夹角的果树点云配准与枝叶分割方法研究

Fruit Tree Point Cloud Registration Based on Normal Vector Angles and Branch-Leaf Segmentation Method

摘要:

在实现果园作业全自动化的过程中，亟需直接构建自然环境下果树枝干三维模型的方法。本文通过对自然环境下以不同角度采集的果树点云进行配准，并针对采样一致性（SAC-IA）+迭代最近点（ICP）配准算法在点云配准中耗时较长以及精度不高的问题，结合点云法向量夹角提取源点云和目标点云的特征点，并通过点云法向量夹角的余弦值在源点云和目标点云的特征点中查找待匹配点对的方法，提出了一种基于果树点云待匹配点对的改进SAC-IA+ICP点云配准算法；借助最小包围盒划分的分块技术对配准后的果树点云进行分块，然后利用点云的几何特征，对划分的子块进行枝叶粗分割，最后使用欧氏聚类完成枝叶的精细分割。对比实验结果显示，改进后的SAC-IA+ICP算法在平均旋转误差上相较于原始SAC-IA+ICP算法减少85.44%，配准均方根误差相较于原始SAC-IA+ICP算法减少71.74%，配准时间相较于原始SAC-IA+ICP算法减少97.99%；同时，改进后的SAC-IA+ICP算法在平均旋转误差上相较于SAC-IA+NDT算法减少90.38%，配准均方根误差相较于SAC-IA+NDT算法减少85.39%，配准时间相较于SAC-IA+NDT算法减少98.04%。另外，本文采用的枝叶分割算法能够完成枝叶分割，且相较于人工分割其分割准确度可达94.77%。

Abstract:

In realizing full automation of orchard operations, it is urgent to construct a 3D model of fruit tree branches and trunks in the natural environment directly. Point clouds of fruit trees collected from different views in the natural environment were registered. Considering that sampling consistency (SAC-IA)+iterative nearest point (ICP) registration algorithm took a long time and had low accuracy in point cloud registration. Thus, the feature points of the source point cloud and target point cloud were extracted by combining the angle of the normal vector of the point cloud, and then matching point pairs were found in the feature points of the source and target point clouds based on the cosine value of the angle of the normal vector of the point cloud. Using the matching point pairs of fruit tree point clouds, an improved SAC-IA+ICP point cloud registration algorithm was proposed. Further, the registered fruit tree point cloud was partitioned by using the partitioning technology of minimum box partition, and then the branches and leaves of the partitioned sub-blocks were roughed by using the geometric features of the point cloud; finally, the branches and leaves were partitioned by using Euclidean clustering. Compared with the original SAC-IA+ICP algorithm, the average rotation error was reduced by 85.44%, and the registration root mean square error can be reduced by 71.74%, the registration time was reduced by 97.99%. Meantime，compared with the SAC-IA+NDT algorithm, the average rotation error was reduced by 90.38%, and the registration root mean square error can be reduced by 85.39%, the registration time was reduced by 98.04%. The segmentation algorithm can complete the segmentation of branches and leaves, and the accuracy can reach 94.77% compared with manual segmentation.