Robust Lifelong Indoor LiDAR Localization Using the Area Graph

Fujing Xie, Sören Schwertfeger

TL;DR

Researchers have developed AGLoc, a localization system that swaps out fragile, high-maintenance occupancy grid maps for "Area Graphs"—compact, semantic maps based on permanent architectural features like walls and doors. By focusing on the building's skeleton rather than transient clutter (like furniture or people), AGLoc achieves robust, lifelong indoor localization that doesn't break when the environment changes.

What Problem Does This Paper Address?

Most autonomous robots navigate using 2D occupancy grids or 3D point clouds. While precise, these maps are "brittle": they capture every piece of furniture, box, and partition. When someone moves a desk or opens a door, the map becomes outdated, leading to localization failure unless the map is constantly updated.

Furthermore, creating these maps usually requires a robot to perform a dedicated SLAM (Simultaneous Localization and Mapping) run. The authors argue that for "lifelong" autonomy, we should instead use maps that are compact (easy to store/transfer), stable (don't require updates for years), and easy to acquire (generated from existing architectural CAD data).

Methods

The authors introduce AGLoc, which utilizes a hierarchical, topometric semantic map called an Area Graph.

Key technical innovations include:

  • Clutter Removal Subsampling: Instead of using the full 3D LiDAR point cloud, the system projects points to 2D and keeps only the furthest points in each angular column. The logic is simple but effective: the furthest points are most likely to be permanent walls rather than nearby clutter like chairs or people.
  • Global Localization via Sampling: To solve the "kidnapped robot" problem, AGLoc uses a rough initial guess from WiFi and barometers (to determine the floor). It then samples thousands of potential poses and scores them against the Area Graph polygons using a custom error function.
  • Weighted Point-to-Line ICP: For active tracking, the system uses an Iterative Closest Point (ICP) variant. It employs a specialized weight function that ignores points "inside" rooms (likely clutter) and points too far away (likely reflections from glass).
  • Corridorness Downsampling: In long, featureless corridors, ICP often "slides" because there are no features to fix the robot's position along the corridor's length. AGLoc detects this "corridorness" using a histogram of wall orientations and downsamples points on the dominant walls to give more weight to rare, non-parallel features (like a door indentation or a distant wall).

Datasets

The authors evaluated their approach using two primary datasets:

  • Seq01: Captured in a heavily cluttered laboratory and a long corridor.
  • Seq02: Captured in an office area and another long corridor.
  • Ground Truth: Generated using LIO-SAM, a high-accuracy SLAM algorithm.
  • Map Source: The Area Graphs were stored in a custom osmAG format (based on OpenStreetMap XML).

Key Insights and Ideas

  • Architecture > Occupancy: By matching against polygons (walls) rather than grids, the robot ignores the "noise" of daily life. A room remains the same room to the robot, even if all the furniture is replaced.
  • The "Furthest Point" Heuristic: This is a clever way to turn a 3D liability (clutter) into a 2D asset. By assuming the furthest return is the wall, they drastically simplify the data processing needed for localization.
  • Solving the Corridor Slide: The "Corridorness" score is a proactive way to handle the classic "degeneracy" problem in ICP. Instead of just failing, the algorithm mathematically rebalances the importance of different sensor readings to find a fix.

Key Results

  • Global Localization: The system achieved a 98-99% success rate in finding the robot's position within 1.0 meter, starting from a broad 6-meter WiFi-based search radius.
  • Efficiency: Pose tracking is incredibly fast, taking only 12.5ms per frame, making it suitable for real-time deployment on modest hardware (tested on an i5 laptop).
  • Robustness: In head-to-head comparisons, AGLoc outperformed the standard AMCL (Adaptive Monte Carlo Localization) when AMCL was forced to use the same sparse Area Graph map, particularly in corridor environments.
  • Compactness: The map representation is orders of magnitude smaller than traditional point clouds or occupancy grids.

Limitations and Future Work

  • Global Localization Speed: While accurate, the global initialization takes about 88 seconds. While acceptable for a one-time startup, the authors aim to optimize this.
  • Sensor Requirements: The current implementation relies on dense, 360-degree LiDAR. It may struggle with older sensors with fewer beams or narrower fields of view.
  • Topological Integration: The current localization mainly uses the metric data (geometry). The authors plan to better integrate the topological "graph" data (how rooms connect) to further improve robustness.
  • Dynamic Doors: While the system handles doors, treating all doors as "closed" or "open" during global localization can still cause slight errors; more sophisticated semantic state estimation for doors is a suggested next step.

Summary updated

Abstract

Lifelong indoor localization in a given map is the basis for navigation of autonomous mobile robots. In this letter, we address the problem of robust localization in cluttered indoor environments like office spaces and corridors using 3D LiDAR point clouds in a given Area Graph, which is a hierarchical, topometric semantic map representation that uses polygons to demark areas such as rooms, corridors or buildings. This representation is very compact, can represent different floors of buildings through its hierarchy and provides semantic information that helps with localization, like poses of doors and glass. In contrast to this, commonly used map representations, such as occupancy grid maps or point clouds, lack these features and require frequent updates in response to environmental changes (e.g. moved furniture), unlike our approach, AGLoc, which matches against lifelong architectural features such as walls and doors. For that we apply filtering to remove clutter from the 3D input point cloud and then employ further scoring and weight functions for localization. Given a broad initial guess from WiFi and barometer localization, our experiments show that our global localization and the weighted point to line ICP pose tracking perform very well, even when compared to localization and SLAM algorithms that use the current, feature-rich cluttered map for localization.
BibTeX
@article{xie2024robust,
  author = {Fujing Xie and Sören Schwertfeger},
  title = {{Robust Lifelong Indoor LiDAR Localization Using the Area Graph}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2024},
  url = {https://openreview.net/forum?id=xNP3NxFZFK},
}

Loading related content...

Figure