A Novel 3D Point Cloud Registration Algorithm Based on Hybrid Line Features

Abstract

Point cloud registration, an approach to recovering the relative transformation of two point clouds, is an essential technique that can be achieved to achieve 3D reconstruction. However, most existing methods are mainly based on point-level features instead of geometric features. These features like lines and planes can be used to intuitively describe the environment and are more reliable than point-level features. Accordingly, this paper proposes an effective registration method based on hybrid line features. The proposed method is constructed in three steps. The first one is the extraction of line features. Inspired by the idea of seeded region growing in image processing, we extract the preliminary line features and then describe them with hybrid descriptors. In the second step, the correspondences of the lines are established using the descriptors. The 2D transformation is then calculated by the candidate correspondences, which registers the point clouds in 2D space to minimize the registration error. Finally, the vertical offset of the point clouds is obtained using the method which is based on the clustering method in the overlapped area, thus lifting the 2D transformation into the final 3D transformation. The experimental results tested on two different kinds of datasets illustrate that the proposed method is effective in achieving high-precision registration results with few line features.