A Fast and Robust 2D LiDAR Alignment Method by Motion Decoupling

Aligning the 2D LiDAR point clouds and estimating their relative poses are the fundamental components of the 2D robotic SLAM (Simultaneous Localization and Mapping) system. Although there are lots of works addressing this problem, existing methods often fail to get a balance between robustness and efficiency. Methods depending on local optimization run fast while sometimes get stuck in local optimum resulting in the wrong pose estimations, which weaken the robustness of the methods. Additionally, methods depending on global optimization always return globally optimal poses by searching the entire given domain for a long period of time, which sacrifice the efficiency of the methods. In this work, we propose a decoupled and globally optimal 2D LiDAR aligning method, which differs from existing methods by achieving the robustness and efficiency of the 2D LiDAR pose estimation simultaneously. Concretely, we use the invariant vectors features decoupling the motion of rotation and translation to reduce the dimensionality of the alignment problem. Consequently, the branch-and-bound algorithm runs in a low dimension to obtain a robust relative pose. Moreover, we propose a feature-selecting strategy to speed up our method. The proposed method is verified in various synthetic and real-world data and shows great performance.