Title:AC-LIO: Towards Asymptotic and Consistent Convergence in LiDAR-Inertial Odometry

Abstract:Existing LiDAR-Inertial Odometry (LIO) frameworks typically utilize prior state trajectories derived from IMU integration to compensate for the motion distortion within LiDAR frames, and demonstrate outstanding accuracy and stability in regular low-speed and smooth scenes. However, in high-speed or intense motion scenarios, the residual distortion may increase due to the limitation of IMU's accuracy and frequency, which will degrade the consistency between the LiDAR frame with its represented geometric environment, leading pointcloud registration to fall into local optima and consequently increasing the drift in long-time and large-scale localization.
To address the issue, we propose a novel asymptotically and consistently converging LIO framework called AC-LIO. First, during the iterative state estimation, we backwards propagate the update term based on the prior state chain, and asymptotically compensate the residual distortion before next iteration. Second, considering the weak correlation between the initial error and motion distortion of current frame, we propose a convergence criteria based on pointcloud constraints to control the back propagation. The approach of guiding the asymptotic distortion compensation based on convergence criteria can promote the consistent convergence of pointcloud registration and increase the accuracy and robustness of LIO. Experiments show that our AC-LIO framework, compared to other state-of-the-art frameworks, effectively promotes consistent convergence in state estimation and further improves the accuracy of long-time and large-scale localization and mapping.