Radar sensors have become increasingly attractive for autonomous mobility systems in adverse weather and low-visibility conditions. Radar sensors are robust to environmental challenges and directly measure relative velocity—two distinct advantages over cameras or light detection and ranging (LiDAR) approaches. Despite these advantages, radar point clouds remain inherently sparse and noisy, typically comprising only a few hundred points per frame; this fact poses considerable challenges for reliable simultaneous localization and mapping (SLAM). Most point cloud SLAM systems are developed for dense LiDAR measurements that are not suitable for radar measurements. In this study, we propose a feature extraction and registration framework designed specifically for radar sensors. Based on LiDAR odometry and mapping (LOAM), we developed feature extraction logic that can handle sparse radar observations more effectively. Using a curvature calculation technique that is robust against noise in radar observations and a principal-component-analysis-based linearity determination technique, we analyze the geometric characteristics of point clouds and identify meaningful points. Furthermore, we utilize Doppler velocity and meta-information to apply radar characteristics and remove dynamic objects and noise that can degrade point-matching performance. These selected features are then used in a two-dimensional iterative closest point alignment method for pose estimation, ultimately enabling robust localization using only sparse point clouds. We validated this logic based on various scenarios using the nuScenes dataset and confirmed that this logic outperforms existing LOAM-based point cloud processing methods. This demonstrates the potential of the proposed method for assisted driving in autonomous mobility systems under extreme situations.
Kang et al. (Tue,) studied this question.