계층적 정합을 통한 KISS-ICP 기반 LiDAR 오도메트리 성능 향상

초록

Light detection and ranging (LiDAR) odometry estimates motion by registering sequential point clouds and constitutes a key component of autonomous navigation systems. Keep it Small and Simple Iterative Closest Point (KISS-ICP) is a recent LiDAR-only odometry method that demonstrates that accurate and robust performance can be achieved using a simple point-to-point ICP formulation without feature extraction or additional sensor fusion. While this simplicity enables efficient and reliable operation in many scenarios, additional improvements in registration stability are possible when large initial pose uncertainty or abrupt motion occurs. This study presents a coarse-to-fine registration approach that complements the original KISS-ICP framework by introducing a hierarchical local map structure with multiple resolutions. In the proposed approach, a low-resolution map is first used to obtain a stable coarse alignment, followed by fine registration against a standard-resolution map to refine the pose estimate. This hierarchical strategy enhances the convergence behavior while preserving the lightweight design philosophy of KISS-ICP. Experiments on the M2DGR and NTU-VIRAL datasets show that the proposed method consistently improves accuracy compared with the original KISS-ICP, particularly in scenarios involving rapid motion and substantial rotational changes. Notably, the proposed method achieves up to 94.89% reduction in vertical endpoint translation error on the M2DGR dataset. These results indicate that the hierarchical registration approach based on the multi-resolution map structure is less sensitive to outliers and initial pose errors than the original KISS-ICP, leading to improved localization accuracy.

키워드