Efficient and Robust Line-based Registration Algorithm for Robot Perception Under Large-scale Structural Scenes

Abstract

Point cloud registration is a classical problem in advanced robot perception. Despite having been widely studied, the registration of large-scale point clouds still remains challenging in terms of both efficiency and accuracy. In this paper, aiming at the registration in large-scale structural scenes that contains numerous line-features, we propose a line-based efficient and robust registration algorithm for robot perception. Concretely, we first extract lines from point clouds and use the line-features to perform the registration, which decreases the scale of algorithm’s input and decouples the rotation and the translation sub-problems. Consequently, it reduces the complexity of registration problem. We then solve the rotation and translation sub-problems using the branch-and-bound algorithm, which ensures the accuracy and robustness of registration. In translation sub-problem, we propose two strategies to adapt to the registration problem in different scenes, the one is universal algorithm, the other is decoupled algorithm. Extensive experiments are performed on both synthetic and real-world data to demonstrate the advantages of our method.