Registration of Noisy Point Clouds Using Virtual Interest Points

A new method is presented for robustly and efficiently registering two noisy point clouds. The registration is driven by establishing correspondences of virtual interest points, which do not exist in the original point cloud data, and redefined by the intersection of parametric surfaces extracted from the data. Parametric surfaces, such as planes, exist in abundance in both natural and artificial scenes, and can lead to regions in the data of relatively low noise. This in turn leads to repeatable virtual interest points, with stable locations across overlapping images. Experiments were run using virtual interest points defined by the intersection of three implicit planes, applied to data sets of four environments comprising100 point clouds. The proposed method outperformed the Iterative Closest Point, Generalized Iterative Closest Point, and a 2.5D SIFT-based RANSAC method in registering overlapping images with a higher success rate, and more efficiently.