SLIM: Scalable and Lightweight LiDAR Mapping in Urban Environments

TL;DR

SLIM tackles the memory and scalability bottlenecks of dense LiDAR maps for long-term urban mapping by representing geometry with lightweight lines and planes. The system merges multi-session maps without heavy place-recognition, and refines them via a two-stage backend: Pose Graph Optimization followed by Bundle Adjustment, aided by a map-centric nonlinear factor recovery to bound pose growth. Extensive experiments on KITTI, NCLT, HeLiPR, and M2DGR show accurate mapping, strong memory efficiency (~$130\, 0\mathrm{KB/km}$ on KITTI), and reliable map reuse for online localization. The work demonstrates that carefully designed front-end representations and a sparse, map-centric back-end can achieve globally consistent urban maps at a fraction of the memory required by traditional point-cloud maps, enabling practical long-term deployment.

Abstract

LiDAR point cloud maps are extensively utilized on roads for robot navigation due to their high consistency. However, dense point clouds face challenges of high memory consumption and reduced maintainability for long-term operations. In this study, we introduce SLIM, a scalable and lightweight mapping system for long-term LiDAR mapping in urban environments. The system begins by parameterizing structural point clouds into lines and planes. These lightweight and structural representations meet the requirements of map merging, pose graph optimization, and bundle adjustment, ensuring incremental management and local consistency. For long-term operations, a map-centric nonlinear factor recovery method is designed to sparsify poses while preserving mapping accuracy. We validate the SLIM system with multi-session real-world LiDAR data from classical LiDAR mapping datasets, including KITTI, NCLT, HeLiPR and M2DGR. The experiments demonstrate its capabilities in mapping accuracy, lightweightness, and scalability. Map re-use is also verified through map-based robot localization. Finally, with multi-session LiDAR data, the SLIM system provides a globally consistent map with low memory consumption (~130 KB/km on KITTI).

Figures (17)

• Figure 1: Demonstration of SLIM using the HeLiPR dataset jung2023helipr. Best viewed zoomed in and in color. SLIM provides parameterized maps with lines (colored in cyan) and planes (colored in magenta). For comparison, we also present point clouds (in light blue), which are widely used in conventional LiDAR mapping systems. More mapping results are presented in Figure \ref{['fig:mapping_visualization']}. Furthermore, we compare the memory consumption of SLIM-generated maps with point cloud maps of varying densities (downsampled with radius $r$). In this HeLiPR session, SLIM maps are inherently more lightweight than LiDAR point cloud maps, with memory usage significantly reduced from 690.60 MB (point clouds) to just 0.41 MB (SLIM).
• Figure 2: System Overview. The front-end map vectorization module extracts features and parameterizes planes to lanes. Then, the SLIM system generates one global lightweight map via map merge. The map refinement modules, including PGO and BA, smooth the mapping results only using lines and planes. The map-centric marginalization can bound the computational and storage requirements with increasing map sessions. Overall, SLIM is a centralized server that can achieve map vectorization, map merge and refinement for long-term LiDAR mapping.
• Figure 3: Formulation of line and plane landmarks. On the left side, the line $\mathbf{l}(x, y)$ can be obtained through a translation $(x, y)$, followed by the derivation of $\mathbf{l}(\mathbf{R}(\alpha, \beta), x, y)$ using a 2-DoF rotation $\mathbf{R}(\alpha, \beta)$. Similarly, on the right side, the plane $\mathbf{s}(d)$ is formulated by applying a translation $d$, and subsequently transformed into $\mathbf{s}(\mathbf{R}(\alpha, \beta), d)$ via rotation. A corresponding transformation always exists for any line or plane to formulate landmarks.
• Figure 4: Two registration cases at different places, using the proposed block registration method. We reduce the number of plane landmarks by clustering, thus speeding up the registration process for map merge. The lines between the clustered landmarks represent the data associations (correspondences) between the block in different coordinates.
• Figure 5: Factor graph of the proposed LiDAR BA. Residuals are constructed from the odometric poses, point-to-line, and point-to-plane measurements. A prior pose factor is utilized to fix the global pose for optimization.