Abstract: 3D point cloud registration is the fundamental of a large number of applications in computer vision, computer graphics, and remote sensing, etc. However, the existence of noises and a low overlapping ratio in the scanning data poses a great challenge to the existing registration methods. Facing the point clouds of man-made objects or urban scenes that are likely to have the aforementioned issues, we propose a registration method of 3D point clouds via matching multi types of geometric primitives. Our method first extracts common geometric primitives from raw point clouds and further builds the feature descriptors from their effective combinations. Then, under the matching of the descriptors, our method realizes the matching of primitives and acquires the transformation parameters from them. Finally, based on the global evaluation for every candidate transformation, the best transformation is identified and applied to achieve the registration. Our method fully inherits the advantages of multiple types of primitives and has stronger robustness and efficiency. Experiments on various benchmarks demonstrate that our method achieves state-of-the-art registration performance.

Key words: 3D point cloud, point cloud registration, geometric primitive, effective combination