PRISM-TopoMap: Online Topological Mapping with Place Recognition and Scan Matching

Kirill Muravyev, Alexander Melekhin, Dmitriy Yudin, Konstantin S. Yakovlev

Abstract

PRISM-TopoMap is a novel online topological mapping system designed for mobile robots. It addresses the limitations of traditional metric maps—such as high memory consumption and sensitivity to odometry drift—by constructing a lightweight graph of locations. The system fuses multimodal place recognition (RGB images and LiDAR) with a geometric scan-matching filter to ensure topological consistency. Experimental results demonstrate that PRISM-TopoMap is significantly more memory-efficient and robust than state-of-the-art metric and topological SLAM methods, maintaining connectivity even under significant sensor noise.

Introduction

Most autonomous robots navigate using metric maps, such as occupancy grids or dense point clouds. While precise, these maps are "heavy," consuming massive amounts of RAM and disk space as the environment grows. Furthermore, they are highly sensitive to odometry drift; small sensor errors accumulate over time, causing the map to warp or fail during loop closures.

Topological mapping offers an alternative by representing the world as a graph of connected "locations" (e.g., rooms or corridor segments). However, existing topological methods often suffer from perceptual aliasing—mistaking two similar-looking but distant places for the same one. PRISM-TopoMap aims to bridge this gap by creating a system that is as lightweight as a graph but as geometrically reliable as a metric map.

Overview of the PRISM-TopoMap system, which converts multimodal sensor data into a sparse topological graph.

Methodology

The PRISM-TopoMap system ensures an accurate and connected topological graph through a three-stage process involving multimodal recognition, geometric verification, and online maintenance.

Multimodal Place Recognition (MSSPlace-G)

The core of the system is a modified version of the MSSPlace model. It processes front-view and back-view RGB images alongside a voxelized LiDAR point cloud. These inputs are passed through 2D and 3D convolutional neural networks (CNNs) and fused into a single "place descriptor." This multimodal approach allows the robot to identify previously visited locations even if visual appearance changes, provided the 3D structure remains consistent.

Architecture of the MSSPlace-G multimodal place recognition model showing the fusion of image and point cloud features.

Geometric Verification and Scan Matching

To prevent the "false positive" matches common in pure learning-based systems, PRISM-TopoMap implements a geometric sanity check. The system projects 3D LiDAR scans into 2D black-and-white images and uses classical computer vision techniques (ORB features) to verify matches. If the geometry does not align, the match is rejected. This step also allows the system to estimate precise relative poses between locations without requiring a global coordinate system.

Online Graph Maintenance

The map is built dynamically. At each step, the system determines if the robot is still within its current node. If the robot moves beyond the current node's boundaries, the system checks neighboring nodes, attempts loop closure via place recognition, and only creates a new node if the robot has entered uncharted territory.

Experimental Results

The system was evaluated using high-fidelity simulations (Matterport3D, Gibson, HM3D) and real-world data from a Husky A200 robot.

  • Efficiency: PRISM-TopoMap demonstrated extreme memory efficiency, using only 0.15 GB of RAM compared to 2.20 GB for Hydra. Its disk footprint was a mere 0.4 MB, whereas ORB-SLAM3’s map reached 237 MB for the same area.
  • Robustness: Under high-noise conditions, PRISM-TopoMap maintained a perfectly connected graph (1.0 connected components), while competing topological methods like IncrementalTopo fragmented into over 17 disconnected pieces.
  • Accuracy: The MSSPlace-G model achieved a 95.26% Recall@1 on the test set, outperforming other multimodal models like MinkLoc++.

Conclusion

PRISM-TopoMap provides a robust solution for long-term robot autonomy by abandoning the requirement for a global coordinate frame in favor of local connectivity. By using geometry as a filter for multimodal place recognition, the system remains immune to the long-term drift that typically breaks metric SLAM systems. Future work will focus on incorporating semantic labeling and integrating specialized path-planning algorithms to leverage these sparse topological structures for faster long-range navigation.

Summary updated

BibTeX
@article{muravyev2025prism,
  author = {Kirill Muravyev and Alexander Melekhin and Dmitriy Yudin and Konstantin S. Yakovlev},
  title = {{PRISM-TopoMap: Online Topological Mapping with Place Recognition and Scan Matching}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2025},
  doi = {10.1109/lra.2025.3541454},
}
Figure