Geometric Multi-Session Map Merging with Learned Local Descriptors

TL;DR

This work tackles large-scale, multi-session LiDAR map merging by introducing GMLD, a learning-based framework that combines a KPConv-based keypoint-aware local descriptor with a plane-based geometrical transformer to capture geometric context. Loop closures are detected and registered through descriptor matching and refined with GICP, while a scan-matching cost factor is integrated into a factor-graph optimization to enforce global geometric consistency across sessions. Key contributions include a geometry-aware downsampling and keypoint detection module, a plane-based transformer encoder, and the use of inter-session scan matching costs within a robust PCM-based outlier rejection pipeline, all evaluated across diverse datasets with strong results. The approach demonstrates accurate, robust, and scalable map merging suitable for online, large-scale autonomous operations with real-time potential.

Abstract

Multi-session map merging is crucial for extended autonomous operations in large-scale environments. In this paper, we present GMLD, a learning-based local descriptor framework for large-scale multi-session point cloud map merging that systematically aligns maps collected across different sessions with overlapping regions. The proposed framework employs a keypoint-aware encoder and a plane-based geometric transformer to extract discriminative features for loop closure detection and relative pose estimation. To further improve global consistency, we include inter-session scan matching cost factors in the factor-graph optimization stage. We evaluate our framework on the public datasets, as well as self-collected data from diverse environments. The results show accurate and robust map merging with low error, and the learned features deliver strong performance in both loop closure detection and relative pose estimation.

Figures (11)

• Figure 1: GMLD provides a robust and precise map merging framework by detecting and registering point clouds in overlapping regions through the utilization of learned discriminative descriptors and scan matching cost factors. The example shown illustrates the accurate integration of the quad and parkland with an additional corridor map from the Newer College.
• Figure 2: The proposed map merging framework consists of three modules: (i) Keypoints and Descriptors Generation extracts keypooints and local descriptors from dense point clouds; (ii) Loop Closure Detection and Registration identifies potential loop closures by computing the average of the smallest $n$ pairwise distances between local descriptors and estimates the relative transformation by aligning the corresponding keypoints; and (iii) Map Merging constructs the pose and scan matching cost factor graph using candidates that satisfy the inlier ratio, relative error, and pairwise consistency maximization checks.
• Figure 3: The neural network comprises two modules: (i) a Keypoints-Aware Downsampling module that identifies consistent keypoints and their associated local descriptors, and (ii) a plane-based geometric self-attention module that enhances the descriptiveness of the extracted local features.
• Figure 4: Each session’s pose graph, containing intra-map odometry factors (black), is serialized. When an inter-map loop closure is detected between the current and previous sessions, a loop closure factor (green) is added based on the relative pose. Scan matching cost factors (orange) are then inserted between overlapping keyframes across sessions—both in explicit overlap regions near the loop closure and implicit overlaps resulting from the loop closure constraint.
• Figure 5: Map merging results with the proposed algorithm GMLD. The first row shows the unaligned sessions from KITTI $00$ and $02$; ARL Mout site; UCLA Engineering IV and Bunche Hall, while the second row presents the corresponding merged maps.