Point Cloud Registration Algorithm Fusing PCA and NDT

Abstract

To solve the problem that the parameter setting of the NDT point cloud registration algorithm requires considerable experience and it easily falls into local extremum when the initial poses between the point cloud to be registered with the target point cloud differ greatly, the point cloud registration algorithm fusing PCA and NDT is proposed. Firstly, PCA method is used to calculate the principal axis directions of two groups of point clouds, and the initial rigid body transformation matrix is acquired. Then, the problem of principal axis inversion is corrected by using the minimum condition of Euclidean distance mean square error when the principal axis of two groups of point clouds are in the same direction. The correct initial transformation matrix after the principal axis correction is calculated, and the coarse registration of point clouds is completed. To reduce the number of processing point, the two groups of point clouds are sampled under the grid. The rigid body transformation matrix acquired by rough registration is used as the initial position and orientation of the NDT registration algorithm for the point cloud precise registration to obtain the final rigid body transformation matrix. The registration experiment was carried out using the Stanford public database point cloud. The result of this experiment shows that the rotation angle error of the point cloud registration algorithm fusing PCA and NDT reaches 10 ^-3 rad, and the translation component error is better than 0.01 mm. The proposed point cloud registration algorithm can effectively solve the initial position and attitude greater difference between point clouds, and it also can solve the problem of falling into local extremum during NDT registration. The registration accuracy basically meets the point cloud registration requirements, which can achieve fast and stable registration result with two groups of point clouds under any position and attitude.