A fast and accurate tightly coupled Light Detection and Ranging-inertial odometry based on sparse voxel maps and Gauss-Newton method

Abstract

Abstract This work proposes an optimization approach for the time-consuming parts of Light Detection and Ranging (LiDAR) data processing and IMU-LiDAR data fusion in the LiDAR-inertial odometry (LIO) method. Two key novelties enable faster and more accurate navigation in complex, noisy environments. Firstly, to improve map update and point cloud registration efficiency, we employ a sparse voxel maps with a new update function to construct a local map around the mobile robot and utilize an improved Generalized Iterative Closest Point algorithm based on sparse voxels to achieve LiDAR point clouds association, thereby boosting both map updating and computational speed. Secondly, to enhance real-time accuracy, this paper analyzes the residuals and covariances of both IMU and LiDAR data in a tightly coupled manner, and achieves system state estimation by fusing sensor information through Gauss-Newton method, effectively mitigating localization deviations by appropriately weighting the LiDAR covariances. The performance of our method is evaluated against advanced LIO algorithms using eight open datasets and five self-collected campus datasets. Results show a 24.7–60.1% reduction in average processing time per point cloud frame, along with improved robustness and higher precision motion trajectory estimation in most cluttered and complex indoor and outdoor environments.