A Novel LiDAR-Inertial SLAM with Low Accumulation Error

Existing LiDAR-based SLAM frameworks typically consist of a frontend odometry and a backend global optimization. The frontend performs real-time ego-motion estimation through point cloud registration, while the backend involves graph optimization and pose smoothing at a lower frequency. To further minimize cumulative error, we enhance the LiDAR-IMU SLAM algorithm. The frontend focuses on estimating relative poses within a local map centered on keyframes by a step-by-step optimization approach, while the backend corrects keyframe poses based on global optimality criteria by integrating ground feature information, loop closure, and a priori maps into the factor graph. The real-time pose output is achieved by merging the optimization results from both modules. Experimental validation is conducted using both public and self-collected datasets. Qualitative and quantitative trajectory analysis demonstrates the effectiveness of our method in reducing cumulative error and ensuring robustness. Furthermore, our algorithm facilitates the expansion of point cloud maps. By partitioning the scene into partially overlapping sub-regions, we can gradually construct large-scale point cloud maps.