LIO-GVM: an Accurate, Tightly-Coupled Lidar-Inertial Odometry with Gaussian Voxel Map

TL;DR

The paper tackles robust LiDAR-IMU odometry by tightly integrating sensor data through an iterated error-state Kalman filter. It introduces Gaussian voxel representations for both correspondences and a lightweight, incrementally updatable voxel map, guided by a new residual that accounts for both distance and variance disparity via distribution-to-distribution matching. Empirical results on diverse datasets show improved accuracy and stability over state-of-the-art methods, along with lower memory usage and competitive timing. The approach is open-sourced, offering a practical, scalable solution for real-time LIO applications.

Abstract

This letter presents an accurate and robust Lidar Inertial Odometry framework. We fuse LiDAR scans with IMU data using a tightly-coupled iterative error state Kalman filter for robust and fast localization. To achieve robust correspondence matching, we represent the points as a set of Gaussian distributions and evaluate the divergence in variance for outlier rejection. Based on the fitted distributions, a new residual metric is proposed for the filter-based Lidar inertial odometry, which demonstrates an improvement from merely quantifying distance to incorporating variance disparity, further enriching the comprehensiveness and accuracy of the residual metric. Due to the strategic design of the residual metric, we propose a simple yet effective voxel-solely mapping scheme, which only necessities the maintenance of one centroid and one covariance matrix for each voxel. Experiments on different datasets demonstrate the robustness and accuracy of our framework for various data inputs and environments. To the benefit of the robotics society, we open source the code at https://github.com/Ji1Xingyu/lio_gvm.

Figures (8)

• Figure 1: The top image illustrates the global map built by LIO-GVM aligned with the Google map. The bottom image shows the runtime voxel map of LIO-GVM inside the rectangle. For better visualization, we model the distributions within each voxel as surfels. The blue, red, green, and white surfels denote the ground, wall, pole, and other types of voxels respectively.
• Figure 2: System Workflow.
• Figure 3: Directly registering the source points to the nearest target points would generate mismatches (shown in the left image). Instead, we: (a). Voxelize the target cloud and fit the points within each voxel with a Gaussian distribution. (b). Fit each source point and its nearest $y$ neighbors with a Gaussian distribution and match it with the near & similar target distributions.
• Figure 4: The different neighbor set $\mathcal{V}_j$ of a given voxel ${}^G\mathbf{v}_j^n$.
• Figure 5: The position error of different methods on sequence ntu_13. We find that the position error is most significant along the z axis, due to the severe elevation changes.