Point Cloud Registration with Low Overlap Based on Dimension Reduction and Feature Matching

Abstract

In the process of point cloud registration with low overlap, existing algorithms have a series of problems of computational complexity and time cost. In order to improve the speed and precision of registration, this paper proposed a method based on dimension reduction and feature matching. Firstly, 3d disordered points in point cloud are transformed into 2d images by projection, where the principal component analysis (PCA) is used to calculate the original projection direction and the perception hash algorithm is used to correct it; Secondly feature extraction and matching are completed in the image space, which sharply decrease the time cost. And unreliable matches are filtered out based on voting rules and distance invariance of the rigid body transformation, which ensure the precision of registration. Finally, transformation matrix is calculated according to the quaternion method, and ICP is used to iteratively calculate the optimization results. The public point cloud data and physical model point cloud data are used as samples in experiments, and the experimental results show that the proposed method can achieve a highly precise registration with an overlap rate as low as 30%. The time cost is reduced by more than 160% compared with the existing algorithms, and speed superiority of the proposed method is more significant with the increase of data scale.