Point cloud registration by combining shape and intensity contexts

Co-registration of point clouds is critical when a scene is measured several times. We present a novel feature based solution, where features are described by combining local shape context and local intensity (or image) context. The Euclidean distances of such shape and intensity combined context descriptors are used to identify candidate correspondences, which are then used as input to the final (co-)registration with the random sample consensus method. The proposed approach is embedded in the concepts of point cloud pyramid and local shape context. The use of pyramid allows for robustness to noise, whereas the new feature descriptor takes into account both local geometry and reflectivity. Tests with three point clouds: one airborne, one terrestrial, and one image derived, demonstrate an automated and robust registration at a precision of 1.4-1.7 times the point spacing, which is sufficient for many applications and can lead to a convergence for further refinement.