MULO: LiDAR Odometry via MUlti-Landmarks Joint Optimization

Abstract

At the present stage, LiDAR-based SLAM solutions are dominated by ICP and its variants, while the BA optimization method that can improve the pose consistency has received little attention. Therefore, we propose MULO, a low-drift and robust LiDAR odometry using BA optimization with plane and cylinder landmarks. In the front-end, a coarse-to-fine direct pose estimation method provides the prior pose to the back-end. And in the back-end, we propose a novel three-stage landmark extraction and data association strategy for plane and cylinder, which is robust and efficient. Meanwhile, a stable minimum parameterization method for cylinder landmarks is proposed for optimization. In order to fully utilize the LiDAR information at long distances, we propose a new sliding window structure consisting of a TinyWindow and a SuperWindow. Finally, we jointly optimize the two kinds of landmarks and scan poses in this sliding window. The proposed system is evaluated on public dataset and our dataset, and experimental results show that our system is competitive compared with the state-of-the-art LiDAR odometrys.