Towards high-performance image-grade LiDAR odometry and mapping

Abstract

Image-grade LiDARs can provide dense range measurements and high angular resolution, which makes them attractive for vehicle localization and map construction. However, the large number of points and the limited field of view also increase the difficulty of real-time and robust odometry estimation, especially in tunnels, urban roads, and other geometrically degraded scenes. In this paper, we develop a LiDAR-only SLAM framework for image-grade LiDARs. The system is formulated atop a factor graph, where denoised and distortion-free point clouds are projected onto a range image. Ground, planar, and edge features are extracted using a degeneracy-aware feature selection strategy for pose estimation, while line features are further incorporated in structured environments. To improve robustness under the limited FoV, a mixed loop detection strategy combining trajectory and structural information is introduced. Experiments in several realworld scenes show that the proposed system can produce consistent trajectories and maps while maintaining real-time performance.