SW-CLS: Point Cloud Registration Method Based on Sliding Window Collar Line Segments

LiDAR technology is widely used in autonomous driving due to its high resolution and long-range detection. However, large-scale point cloud data require significant computational resources, limiting its application. This article addresses the challenge of computing transformations between sparse point clouds in LiDAR odometry. To overcome issues related to point cloud sparsity, ring-like structures, and discontinuous feature loss in Collar Line Segment (CLS) sampling, we propose a Sliding Window CLS (SW-CLS) point cloud registration method. Unlike CLS, which segments point clouds by fixed horizontal angles, SW-CLS uses a SWapproach to extract more continuous and richer line features. Registration is achieved by minimizing the distance between corresponding lines in adjacent point clouds. Experiments on public datasets demonstrate that SW-CLS achieves lower odometry drift than CLS in environments with abundant planar features, reducing absolute trajectory errors (ATEs) in rotation and translation by up to 69.5% and 46.5%, respectively, while also improving relative trajectory accuracy.