MS-Mapping: An Uncertainty-Aware Large-Scale Multi-Session LiDAR Mapping System

TL;DR

This work tackles the challenge of large-scale, multi-session LiDAR mapping by introducing MS-Mapping, an uncertainty-aware, incremental system that reduces data redundancy through distribution-aware keyframe selection and improves back-end robustness with an uncertainty-driven Pose-SLAM framework. The core innovations include a Wasserstein-distance-based map-distribution measure for real-time keyframe selection, an IEKF-inspired uncertainty model that adapts constraints without scene-specific tuning, and a comprehensive evaluation benchmark with ground-truth maps across indoor and outdoor scenarios. Empirical results on public datasets and a large-scale ground-truth map demonstrate superior map accuracy, robustness to loop closures, and coherent global maps compared to state-of-the-art baselines, while also highlighting runtime considerations. The work points to practical impacts in surveying, urban mapping, autonomous navigation, and multi-agent exploration, with potential extensions to place-recognition integration and crowd-sourced, multi-robot mapping.

Abstract

Large-scale multi-session LiDAR mapping is essential for a wide range of applications, including surveying, autonomous driving, crowdsourced mapping, and multi-agent navigation. However, existing approaches often struggle with data redundancy, robustness, and accuracy in complex environments. To address these challenges, we present MS-Mapping, an novel multi-session LiDAR mapping system that employs an incremental mapping scheme for robust and accurate map assembly in large-scale environments. Our approach introduces three key innovations: 1) A distribution-aware keyframe selection method that captures the subtle contributions of each point cloud frame to the map by analyzing the similarity of map distributions. This method effectively reduces data redundancy and pose graph size, while enhancing graph optimization speed; 2) An uncertainty model that automatically performs least-squares adjustments according to the covariance matrix during graph optimization, improving mapping precision, robustness, and flexibility without the need for scene-specific parameter tuning. This uncertainty model enables our system to monitor pose uncertainty and avoid ill-posed optimizations, thereby increasing adaptability to diverse and challenging environments. 3) To ensure fair evaluation, we redesign baseline comparisons and the evaluation benchmark. Direct assessment of map accuracy demonstrates the superiority of the proposed MS-Mapping algorithm compared to state-of-the-art methods. In addition to employing public datasets such as Urban-Nav, FusionPortable, and Newer College, we conducted extensive experiments on such a large \SI{855}{m}$\times$\SI{636}{m} ground truth map, collecting over \SI{20}{km} of indoor and outdoor data across more than ten sequences...

Figures (22)

• Figure 1: Visualization of incremental mapping results using the MS-mapping algorithm with the short-exp11 (S2) and the parkland0 sequence (S1) (Table \ref{['tab:algorithm-comparison-nc']}). (A) Ground truth map. (D) Merged map and trajectory. (B)-(C) Real-world images of regions b and c, where baseline algorithms show significant errors. (b)-(c) Point cloud map of regions a and b in (D).
• Figure 2: Frequent appearance of ill-posed graph optimization in large-scale SLAM. The top subfigure shows the point cloud map of the Mongkok sequencehsu2023hong constructed using the FPGO. Regions a and b illustrate how pose trajectories are distorted during graph optimization with incorrect noise settings in dense loop closure scenarios, which lead to ill-posed optimization and mapping failure. Properly modeled uncertainty, as shown in subfigures A and B, allows the system to effectively optimize these areas.
• Figure 3: Overview of system architecture. Our MS-Mapping system integrates data from distinct agents. The process begins with constructing a pose graph and map frames for the old session (RB2, path: A$\rightarrow$C$\rightarrow$A$\rightarrow$B). Subsequently, we incrementally update the map with data from a new session (PK1, path: A$\rightarrow$C$\rightarrow$D$\rightarrow$A). Key frames are selected based on their contribution to the map using our distribution-aware method. The system performs UPGO to produce a merged pose graph and point cloud map. The right panel illustrates the projection of the final trajectory onto satellite maps.
• Figure 4: Evaluation of discrepancies between point cloud maps $M_1$ and $M_2$ using the Wasserstein distance in (\ref{['eq:wasserstein_distance']}). This metric captures global and local distribution differences. After voxelization, the point count of each voxel denotes its mass. The Wasserstein distances between voxel pairs are averaged to represent the overall difference between GMM maps.
• Figure 5: Incremental update of the GMM map. Each new point cloud frame is transformed into the map based on its pose. The voxels Gaussian parameters are updated point by point, while voxels outside the radius are removed. This process shows the impact of each new frame on the overall map distribution.