Abstract: Aiming at solving the conflicts between convergence region and accuracy of classical registration algorithms, a new variable scale registration algorithm for scattered point clouds is proposed. A measure function used in point clouds registration is constructed on the basis of coincidence point counting. The Gaussian smoothing procedure of the point counting function is investigated, and the relationship between the smoothed measure function and the scale parameter is discussed. The smoothed measure function has the following characteristics：Smooth but extreme-offset using large scale parameter and accurate-global-extreme but local-extreme- existence using small scale parameter. A variable scale registration algorithm for scattered point clouds is proposed on the basis of these characteristics. Inspired by the simulated annealing algorithm, the Lundy annealing strategy is selected as the scale parameter evolution strategy based on comparative experiment. To improve the efficiency of the algorithm, curvature constraint is used to filter the control points; fast Gauss transform is used to speed up the computation of the measure function. At last, the improved convergence region and accuracy of our algorithm are validated through comparison experiments using synthetic and real point clouds.

Key words: Curvature constraint, Fast Gauss transform, Scattered point cloud, Variable scale registration