AGPC-SLAM: Absolute Ground Plane Constrained 3D Lidar SLAM

Abstract

3D lidar-based simultaneous localization and mapping (SLAM) is a well-recognized solution for mapping and localization applications. However, the typical 3D lidar sensor (e.g., Velodyne HDL-32E) only provides a very limited field of view vertically. As a result, the vertical accuracy of pose estimation suf­ fers. This paper aims to alleviate this problem by detecting the absolute ground plane to constrain vertical pose estimation. Different from the conventional rel­ ative plane constraint, this paper employs the absolute plane distance to refine the position in the z-axis and the norm vector of the ground plane to constrain the attitude drift. Finally, relative positioning from lidar odometry, constraint from ground plane detection, and loop closure are integrated under a proposed factor graph-based 3D lidar SLAM framework (AGPC-SLAM). The effectiveness is verified using several data sets collected in Hong Kong.

Problem Addressed

Typical 3D LiDAR sensors provide only a limited vertical field of view, so vertical pose estimation in 3D LiDAR SLAM suffers from accumulated drift; existing relative plane constraints cannot bound this error over time.

Key Contributions

• Absolute ground plane constraint using absolute plane distance and normal vector to constrain z-axis position and attitude drift
• General factor-graph 3D LiDAR SLAM framework integrating LiDAR odometry, ground plane constraint, and loop closure
• Framework readily extensible to additional sensors such as GNSS and magnetometers
• Validation on two challenging large-scale datasets collected in Hong Kong

Conclusions

AGPC-SLAM achieved improved accuracy over the state-of-the-art LeGO-LOAM, effectively mitigating z-axis drift in both evaluated scenarios; better-constrained vertical states also improved x- and y-axis positioning. With loop closure it reached 0.43 m translation RMSE, 0.64 degree rotation RMSE, and 0.13 m altitude error. A limitation is the requirement of a flat ground plane; slopes or ground occlusion by dynamic vehicles can cause misidentification, which future work will address.