LVIG-SAM: LiDAR-Vision-Inertial Odometry with Ground-constraints via Smoothing and Mapping

Existing multi-sensor fusion SLAM systems may achieve precise location and map building in challenging situations. However, for the LiDAR SLAM system, there is still the problem of Z-axis drift due to LiDAR measurement bias. We propose an improved LVIG-SAM based on ground constraints, which can effectively solve the position drift problem caused by LIDAR measurement bias. LVIG-SAM improves on the open-source LVI-SAM by utilizing previously existing ground information. Firstly, the ground points are accurately extracted by identifying and analysing the ground features, and then the RANSAC algorithm is applied to fit these ground points into a ground plane model reflecting the real ground conditions. A Gaussian Newton method is used to optimise the loss function and integrate this optimised information into the factor graph. Compared to the original algorithm, which only uses IMU for z-value estimation, the improved algorithm achieves a more accurate joint estimation by fusing the ground constraints and IMU results. Experimental validation is performed on the M2DGR dataset, which demonstrates more excellent performance compared to the original algorithm.