点云多法向量邻域特征配准算法  

Point cloud registration algorithm based on neighborhood features of multi-scale normal vectors

作　　者：陆军[1] 彭仲涛 夏桂华[1] 

机构地区：[1]哈尔滨工程大学自动化学院,黑龙江哈尔滨150001

出　　处：《光电子．激光》2015年第4期780-787,共8页Journal of Optoelectronics·Laser

基　　金：黑龙江省自然基金(F201123);中央高校基本科研业务费专项基金(HEUCFX41304)资助项目

摘　　要：针对三维激光扫描点云数据的配准问题,提出了一种多法向量邻域特征点云配准算法。首先,根据目标点选取不同邻域半径估算的法向量存在方向偏差,设定约束条件选择关键点,使得初始点云数据量得到精简;其次,设计了一种依据邻域多法向量计算的特征描述子,并计算所有关键点的特征向量;然后,依据所求的特征描述子,使用最小距离与次小距离比值阈值方法初步获取对应关系,并使用随机采样一致性算法和聚类分选方法进行两次优化,得到精确的点与点对应关系;最后,使用奇异值分解法解算刚体变换矩阵,得到配准参数。实验结果表明,由本文设计的关键点选取、特征描述子提取和对应关系筛选方法原理简单、稳定可靠、计算速度较快且计算复杂度小,无需进行第二次配准,对实现点云配准具有实用价值。To solve the registration of 3D laser scanning point cloud data, a new method of registration al- gorithm based on neighborhood features of multi-scale vectors is proposed. Firstly, because there is error between normal vectors of a point calculated by different neighborhood radii, setting constraint condition can be used to select the key points. Thus,the point cloud data is streamlined. Secondly,a method for ex- tracting point feature information is designed based on neighborhood eigenvectors and feature descriptor of all key points can be gotten by using this method. Then, by using the minimum and second distance ratio thresholds to obtain rough corresponding relation and twice optimization methods （random sample consensus algorithm and clustering sorting method）, the exact correspondence between source point and target point cloud can be gotten. Finally,covariance matrix is built and decomposed to get the rigid body transformation matrix. The experimental results show that the selection of key points, extraction of point feature information and determination of correspondence of the new method have simple theory, stable performance, high calculation speed and low computational complexity, and it has practical significance to the realization of point cloud registration.

关 键 词：点云配准 多法向量邻域特征 主成分分析(PCA) 法向量 对应关系 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]