Quatro++: Robust Global Registration Exploiting Ground Segmentation for Loop Closing in LiDAR SLAM

TL;DR

Quatro++ tackles the robustness gap in global LiDAR registration for loop closing by introducing ground segmentation to filter featureless ground points, enabling a decoupled quasi-SE(3) pose estimation that emphasizes yaw-dominant rotation. The framework combines fast ground segmentation, descriptor-based matching, graph-based outlier pruning, and a graduated non-convexity–driven quasi-SO(3) rotation with component-wise translation estimation, all in a learning-free pipeline. Empirical results across KITTI, MulRan, NAVER, and Hilti-Oxford demonstrate higher loop-closure success rates and more accurate SLAM constraints, especially in distant or corridor-like environments, while keeping computation under 1 second. The approach also integrates with INS for roll-pitch compensation, further improving accuracy, and demonstrates tangible benefits when deployed as a back-end loop-closure module in LiDAR SLAM frameworks.

Abstract

Global registration is a fundamental task that estimates the relative pose between two viewpoints of 3D point clouds. However, there are two issues that degrade the performance of global registration in LiDAR SLAM: one is the sparsity issue and the other is degeneracy. The sparsity issue is caused by the sparse characteristics of the 3D point cloud measurements in a mechanically spinning LiDAR sensor. The degeneracy issue sometimes occurs because the outlier-rejection methods reject too many correspondences, leaving less than three inliers. These two issues have become more severe as the pose discrepancy between the two viewpoints of 3D point clouds becomes greater. To tackle these problems, we propose a robust global registration framework, called \textit{Quatro++}. Extending our previous work that solely focused on the global registration itself, we address the robust global registration in terms of the loop closing in LiDAR SLAM. To this end, ground segmentation is exploited to achieve robust global registration. Through the experiments, we demonstrate that our proposed method shows a higher success rate than the state-of-the-art global registration methods, overcoming the sparsity and degeneracy issues. In addition, we show that ground segmentation significantly helps to increase the success rate for the ground vehicles. Finally, we apply our proposed method to the loop closing module in LiDAR SLAM and confirm that the quality of the loop constraints is improved, showing more precise mapping results. Therefore, the experimental evidence corroborated the suitability of our method as an initial alignment in the loop closing. Our code is available at https://quatro-plusplus.github.io.

Figures (23)

• Figure 1: Overview of our global registration method, called Quatro++. As an extension of our previous work lim2022single, Quatro++ consists of five parts. The red, green, and blue lines of the raw and pruned correspondences (shown inside dotted boxes), i.e. $\mathcal{A}_\text{raw}$ and $\mathcal{A}$, respectively, denote the outliers, true inliers, and quasi-inliers (see Section \ref{['sec:rot_estimation']}), respectively. (a) The preprocessing step using ground segmentation (gray points denote the estimated ground points). (b) Correspondence estimation by using descriptor extraction and matching, which can be interchangeable with other methods. (c) Graph-based outlier pruning to initially reject outlier correspondences. (d) Quasi-SO(3) estimation based on graduated non-convexity (GNC). (e) Component-wise translation estimation. (best viewed in color).
• Figure 2: (L-R): Feature matching results between frames 939 as a source cloud (cyan), and 4,206, 4,204 and 4,195 as target clouds (yellow), respectively, for Seq. 02 in the KITTI dataset. The farther the pose discrepancy is, the worse quality of the estimated correspondences is, showing an increase in the ratio of outliers (red lines) and a decrease in the number of actual inliers (green lines).
• Figure 3: (a) Illustration of Patchwork lim2021patchwork. CZM, R-GPF, and GLE are abbreviations of concentric zone model, region-wise ground plane fitting, and ground likelihood estimation, respectively. (b)-(c) Visualization of the ground segmentation output in Seq. 00 in the KITTI dataset for comparison. The green and red colors denote the estimated ground and non-ground points, respectively (best viewed in color). (b) Estimated ground points by using ground segmentation in LeGO-LOAM shan2018lego and (c) by using Patchwork lim2021patchwork. Empirically, it was shown that Patchwork shows more robust segmentation with fewer false positives and false negatives (best viewed in color).
• Figure 4: Visual description of the procedure of Quatro when the frames 2,328 and 3,268 for Seq. 00 in the KITTI dataset are given as source (cyan) and target (yellow), respectively, which is a distant and partially overlapped case. (a) Estimated correspondences, $\mathcal{A}_\text{raw}$, which include potential outliers owing to false matching. The left-top text represents the pose discrepancy. (b) Filtered correspondences by using MCIS-heuristic, which is a graph-based outlier pruning method. (c)-(e) Procedure of Quasi-SO(3) estimation based on graduated non-convexity (GNC). (c) On the initialization step, $\mathcal{A}$ is transformed into translation invariant measurements (TIMs) all of whose weights are set to one (brown color). (d) Next, GNC-based optimization is performed to estimate relative rotation while rejecting outliers by reweighting the weights. (e) The problem is that less than three pairs are occasionally left, which is highlighted by red circles, by assigning near-zero values to the weights of correspondences. Despite this degeneracy, our quasi-SO(3) estimation robustly outputs the relative rotation because its minimum required DoF is one. (f)-(i) Examples of component-wise translation estimation (COTE) to estimate the relative $x$ value. (f) Boundary interval set ${^1\mathcal{E}}$, where the brackets [ and ] denote the lower and upper bounds, respectively. (g) Cardinality of the $g$-th consensus set, $\abs{{^1\mathcal{I}}_g}$. (h) Each value of the optimal solutions, i.e. ${^1\hat{\mathbf{t}}}_g$, which is associated with each ${^1\mathcal{I}}_g$. Note that ${^1\hat{\mathbf{t}}}_{14}$ is not assigned because ${ }^1 \mathcal{I}_{14}=\varnothing$. (i) Residual of the objective function for all correspondences when ${^1\mathbf{t}}_g$ is given. The residual becomes lowest when $g = 8$, thus ${^1 \hat{\mathbf{t}}}_{8}$ is selected as the final solution for the 1st element, i.e. ${^1 \hat{\mathbf{t}}} \leftarrow {^1 \hat{\mathbf{t}}}_{8}$ (best viewed in color).
• Figure 5: (a) Visual description to show the geometrical characteristics of loop closing situations in urban canyons. (a) Revisit usually occurs in either forward, reverse, or perpendicular direction. Accordingly, the roll and pitch differences between the two viewpoints of source and target clouds are sufficiently small. (b) Qualitative analysis of the actual geodesic magnitude of relative roll ($\angle \mathbf{R}_x$), pitch ($\angle \mathbf{R}_y$), and the combined rotations ($\angle (\mathbf{R}_y \cdot \mathbf{R}_x)$) except yaw rotation, where $\angle \mathbf{R}=\cos^{-1}{\frac{\Tr(\mathbf{R}) - 1}{2}}$. The loop pairs whose distances are between 0.5 to 10 m apart in the KITTI dataset were used. (c) Surrogate function $\rho_{\mu}(\cdot)$ when $\bar{c}=0.15$. As the parameter $\mu$ gradually grows, the kernel has more non-linearity. It becomes truncated least squares once $\mu$ reaches $\infty$ (best viewed in color).