关键词: 点云配准, 快速点特征直方图, 随机采样一致, 迭代最近点, KD-Tree

Abstract: To the problem of traditional registration algorithm are difficult to get the desired effect in large-angle pointcloud registration. We proposed a registration method based on exact correspondence feature point pairs and its K-neighbour pointclouds. Firstly, calculate the FPFH of the pointclouds separately, establishing correspondences between point clouds According to the eigenvalues; Then remove the erroneous matching point pairs by RANSAC, and obtain a relatively accurate set of feature point pairs; Moreover, using KD-tree search get the R-Rad region of the feature point pairs respectively, and applying ICP to obtain the optimal convergence of pointclouds. Finally, applying the ICP relative position relationship to the original pointclouds to get the final registration result. Through registration testing and comparison of the Stanford Dragon, Happy Buddha pointcloud models, and Gypsum data scanned by Kinect, The experiment shows that this method can effectively solve the registration problem of pointclouds with large angle transformation, its a 3D pointclouds registration method with high accuracy and robustness.

Key words:  pointcloud registration, fast point feature histograms, random sample consensus, iterative closest point, KD-tree