Local range image descriptor for general point cloud registration

Abstract

The general point cloud registration problem is addressed in this paper. Different from classical descriptors for alignment of RGB-D image, the present work proposes a descriptor for general point clouds, where no lattice structure is defined. The paper considers a point cloud as a set of 3D points without connectivity or normal vectors. In order to construct the descriptor a local reference frame is defined using PCA in the neighborhood of a keypoint and this basis is used to define a local range image upon which the descriptor is constructed. A histogram based approach is used to achieve invariance to rotation. In order to filter out ambiguity and reduce false correspondences a subset of corresponding points was constructed as an instance of the maximum clique problem. The reliability of the proposed descriptor is validated using ROC curve in comparison with other works from literature. Experiments show that the proposed descriptor have high matching precision and that the precision is improved by using the false correspondence filtering process. Furthermore, the false correspondence filtering process proposed in this paper is successfully applied to improve the precision of other descriptors from literature.