ADAPTIVE MOVING LEAST-SQUARES SURFACES FOR MULTIPLE POINT CLOUDS REGISTRATION

摘要

In this paper, we propose an Adaptive Moving Least-Squares (AMLS) surface based approach for multi-view or multi-sensor point cloud ICP registration. The core idea of this approach is to reconstruct a smooth and accurate surface, e. s. AMLS surface, from a point cloud, without data segmentation and surface model selection, resulting in an accurate point-to-AMLS surface ICP registration. The major difference between AMLS and traditional MLS is that the width of Gaussian kernel is adaptively scaled with the principle curvature, which is defined through local integral invariant analysis. Experimental results of both synthetic data and scanned data from a mechanical part show that the presented approach is more accurate and robust on sensor noise and sample density.