A dual quaternion-based, closed-form pairwise registration algorithm for point clouds

摘要

The representation of similarity transformation in three-dimensional (3D) space, especially of orientation, is a crucial issue in navigation, geodesy, photogrammetry, robot arm manipulation, etc. Considering the large amount of computer resources required by iterative algorithms designed for spatial similarity transformation, the high dependence on initial values of unknown parameters, and the instability of solving transformation parameters for large-angle registration, a closed-form solution for pairwise light detection and ranging (LiDAR) point cloud registration is proposed. In this solution, dual-number quaternions are used to represent the 3D rotation. The relationship between the rotation matrix-based representation of similarity transformation and the dual quaternion-based representation is described first. Considering that the same features from two neighboring stations coincide after pairwise registration, a dual quaternion-based error norm, which is associated with the sum of the position errors, is constructed. Based on theory of least squares and by extreme value analysis of the error norm, detailed derivations of the model and the main formulas are obtained. Once the similarities between the same features from the two neighboring LiDAR stations are constructed, the rotation matrix, the scale parameter, and the translation vector are simultaneously derived. Two experiments are conducted to verify the feasibility and effectiveness of the proposed algorithm. The proposed algorithm has the advantages of simplicity and ease of implementation, making it better than the traditional methods that use matrices to describe spatial rotation. Moreover, it solves the transformation parameters without the initial estimates of unknown parameters, making it better than iterative algorithms. Most importantly, in contrast to unit quaternion-based algorithms, the proposed algorithm solves seven unknown parameters simultaneously. Therefore, it effectively avoids the accumulation of introduced error in calculation and the negative impact from the inappropriate choice of initial values.