Automatic registration of large-scale building point clouds with high outlier rates

Abstract

Point cloud registration plays a crucial role in processing large-scale building point cloud data. However, existing registration algorithms face challenges in effectively handling outliers in descriptor-based correspondence. This paper presents an automatic registration method for large-scale building point clouds that is capable of achieving swift and accurate registration without the need for initial guessing. The method employs a two-step matching optimization approach: coarse (two-point)-to-fine (three-point), selecting matches based on two-point reliability and three-point consistency. Spatial transformation parameters are broken down into rotations and translations. A progressively optimized kernel function is proposed for estimating rotation, while a clustering confidence algorithm computes translation. Comprehensive experiments were conducted using real-world data. The results indicate that the approach swiftly and accurately estimates optimal outcomes when processing large-scale building point clouds with outlier rates up to 99%. Compared to six existing registration methods, the proposed approach reduces rotation error by 6.15% and translation error by 12.83%, while improving efficiency by 2.57%.