SCL-SLAM: A Scan Context-enabled LiDAR SLAM Using Factor Graph-Based Optimization

Abstract

In this paper, we present a complete LiDAR SLAM framework, SCL-SLAM, by integrating the loop closure module with the Scan Context descriptor into the tightly-coupled LiDAR-Inertial odometry FAST-LIO2. As a front-end, the direct LiDAR-Inertial odometry module efficiently and robustly produces motion estimates and undistorted scans. Toward the global localization based on 3D LiDAR scans, the lightweight Scan Context descriptor is used in the loop detection module. Additionally, the scan input is filtered through the keyframe selection module to improve the computation efficiency. As a back-end, a pose graph optimization is performed for the optimized trajectory and globally consistent map. SCL-SLAM is extensively evaluated on public datasets and a robot platform over various scales and environments. Experimental result shows that SCL-SLAM achieves higher accuracy than other state-of-art LiDAR SLAM systems and real-time performance. We also extend the proposed system to a centralized architecture SLAM framework for the robot team to use with 3D LiDAR observations.