Multi-Mapcher: Loop Closure Detection-Free Heterogeneous LiDAR Multi-Session SLAM Leveraging Outlier-Robust Registration for Autonomous Vehicles

TL;DR

This paper tackles the challenge of long-term MSS across heterogeneous LiDAR sensors by removing reliance on loop closure detection. It introduces Multi-Mapcher, which performs large-scale map-to-map initial alignment using outlier-robust 3D registration, followed by scan-level refinement and anchor-node based pose graph optimization to fuse multiple sessions into a coherent global map. The approach demonstrates strong robustness to partially overlapped scenes and dynamic changes, while delivering faster inter-session alignment than LCD-based methods across diverse datasets (HeLiPR, HILTI 2021, MulRan, KITTI). The key contributions include (i) LCD-free inter-session initialization via map-to-map and scan-to-scan registration, (ii) anchor-node pose graph optimization for multi-session consistency, (iii) a unified registration pipeline capable of handling heterogeneous LiDAR sensors, and (iv) comprehensive ablations and runtime analysis showing practical efficiency. Overall, Multi-Mapcher advances long-term autonomy by enabling reliable, sensor-agnostic MSS suitable for diverse autonomous platforms.

Abstract

As various 3D light detection and ranging (LiDAR) sensors have been introduced to the market, research on multi-session simultaneous localization and mapping (MSS) using heterogeneous LiDAR sensors has been actively conducted. Existing MSS methods mostly rely on loop closure detection for inter-session alignment; however, the performance of loop closure detection can be potentially degraded owing to the differences in the density and field of view (FoV) of the sensors used in different sessions. In this study, we challenge the existing paradigm that relies heavily on loop detection modules and propose a novel MSS framework, called Multi-Mapcher, that employs large-scale map-to-map registration to perform inter-session initial alignment, which is commonly assumed to be infeasible, by leveraging outlier-robust 3D point cloud registration. Next, after finding inter-session loops by radius search based on the assumption that the inter-session initial alignment is sufficiently precise, anchor node-based robust pose graph optimization is employed to build a consistent global map. As demonstrated in our experiments, our approach shows substantially better MSS performance for various LiDAR sensors used to capture the sessions and is faster than state-of-the-art approaches. Our code is available at https://github.com/url-kaist/multi-mapcher.

Figures (12)

• Figure 1: Multi-session simultaneous localization and mapping (MSS) results of (a) LT-mapper kim2022lt, which is a baseline that relies heavily on a loop closure detection (LCD) module when finding inter-session loop pairs, and (b) our proposed method called, Multi-Mapcher. Gray, dark cyan, and dark magenta colors indicate each session obtained by Ouster OS2-128, Livox Avia, and Aeva Aeries II, respectively. Note that our Multi-Mapcher robustly aligns different sessions from heterogeneous LiDAR sensors while minimizing the dependency on the LCD modules (best viewed in color).
• Figure 2: Overview of our multi-session SLAM framework, called Multi-Mapcher, which consists of four steps (best viewed in color). (a) First, intra-session SLAM is performed using existing SLAM frameworks wei2022fastlio, which outputs scans and optimized poses, including intra-session odometry and loop constraints. (b) Second, map-to-map level registration is performed to initially estimate the relative pose between the reference frames of query session $Q$ and central session $C$. (c) Third, after the initial alignment between two sessions, inter-session loops can be easily detected by a radius search, which is followed by our truncated mean squared error (t-MSE)-based false loop rejection to filter erroneous loop candidates (see Section \ref{['sec:scan_to_scan']}). Note that our Multi-Mapcher uses the same outlier-robust registration module when performing registration at both the map-to-map and scan-to-scan levels. (d) Finally, taking all the odometry, intra-session, and inter-session constraints as inputs, anchor node-based pose graph optimization is performed to build a global map across multiple sessions.
• Figure 3: Original mean squared error (MSE) and our truncated MSE (t-MSE) of submap-to-submap registration results for inter-session loop constraints, where query (from the query session) and target (from the central session) point clouds are acquired by different type of LiDAR sensors. (L-R): The query and target are obtained by Livox Avia (cyan) and Ouster OS2-128 (gray), and by Aeva Aeries II (magenta) and Ouster OS2-128 (gray), respectively (best viewed in color).
• Figure 4: Before and after the application of our anchor node-based pose graph optimization to the trajectory in DCC05 of the HeLiPR dataset jung2023helipr. Note that despite the erroneous trajectory from a single session, our approach successfully minimizes the trajectory errors, which are highlighted as orange dashed boxes (best viewed in color).
• Figure 5: (a)-(c) Qualitative comparison with the state-of-the-art MSS approaches on DCC05, KAIST05, Roundabout03, and Town03 in the HeLiPR dataset (from top to bottom). Gray, dark cyan, and dark magenta colors indicate each session obtained by Ouster OS2-128 (O), Livox Avia (L), and Aeva Aeries II (A), respectively. Black boxes zoom in on specific areas to highlight the misalignment of LT-mapper kim2022lt with STD yuan2023std and to showcase the successful MSS results achieved by our Multi-Mapcher. Red, yellow, and green boxes, outlining subfigures, indicate the failure in both O-A and O-L sessions, failure in at least one of the sessions, and success in both sessions, respectively (best viewed in color).