LT-mapper: A Modular Framework for LiDAR-based Lifelong Mapping

Giseop Kim, Ayoung Kim

TL;DR

LT-mapper is an open-source, modular framework designed for LiDAR-based lifelong mapping in ever-changing urban environments. By breaking the problem into multi-session SLAM, robust change detection, and efficient "delta map" management, it enables robots to maintain up-to-date maps over years while reducing storage costs by up to 60%.

What Problem Does This Paper Address?

Autonomous robots operating in the real world face a major hurdle: the world is not static. Buildings go up, trees are cut down, and cars move. For a robot to navigate reliably over months or years, its map must evolve.

Existing lifelong mapping approaches often struggle with three main issues:

  1. Scalability: Comparing every new scan to a massive global map is computationally expensive.
  2. Alignment Errors: If two maps from different times aren't perfectly aligned, a change detection algorithm might falsely flag a slight shift in a wall as a "new object."
  3. Storage Efficiency: Storing a full 3D point cloud "snapshot" for every single session is unsustainable for long-term deployments.

LT-mapper tackles these by providing a unified, modular pipeline that can handle "permanent" year-level variations (like construction) while remaining computationally and memory efficient.

Methods

The framework is composed of three distinct modules that communicate via a file-based protocol, allowing them to run independently:

  • LT-SLAM: This is the "stitching" engine. It uses a multi-session SLAM (MSS) approach with "anchor nodes" to align temporally distant sessions. Instead of assuming perfect initial alignment, it uses Scan Context for robust loop detection and jointly optimizes trajectories from different times into a shared coordinate system.
  • LT-removert: This module handles change detection. It distinguishes between High Dynamic (HD) objects (like moving cars/pedestrians) and Low Dynamic (LD) changes (like a new construction wall). It uses a "remove-then-revert" algorithm based on range-image visibility, making it robust to the small alignment errors that inevitably occur in real-world SLAM.
  • LT-map: This is the management layer. Instead of saving full maps, it introduces the concept of delta maps—storing only the additions (positive differences) and deletions (negative differences) between sessions.
The modular pipeline of the LT-mapper system, showing the flow from raw data to managed maps.

Datasets

The authors evaluated LT-mapper using two primary sources:

  • MulRan Dataset: Specifically the KAIST sequences, which include data from June 2019 and February 2021. This provides a challenging 1.5-year temporal gap with significant structural changes.
  • LT-ParkingLot Dataset: A new dataset introduced in this paper, consisting of six sessions collected over three days in a parking lot to specifically observe low-dynamic changes (parked cars appearing and disappearing).

Key Insights and Ideas

  • Delta Maps are Game-Changers: By storing only changes rather than full snapshots, the system is incredibly efficient. In one experiment, a delta map required only 0.47 million points compared to 11 million for the full map—a 95.7% reduction in data transmitted to a server.
  • Change Composition: Because the framework is modular and uses delta maps, you can "compose" changes. If you have changes from Session A to B, and B to C, you can mathematically derive the change from A to C without re-running the heavy change detection algorithms. This allows for easy "map rollback" to any point in time.
  • Strong vs. Weak Changes: The authors distinguish between "strong" changes (permanent structures) and "weak" changes (temporary ones like parked cars). This allows the system to generate a Meta Map containing only stable, persistent structures, which is far more reliable for long-term localization than a standard "Live Map."
A visualization of the LT-removert process, showing how it cleans maps and segments changes.

Key Results

  • Robust Alignment: LT-mapper successfully registered maps with a 1.5-year gap, accurately detecting pointwise changes like new construction walls that appeared in the interim.
  • Quantitative Efficiency: Compared to traditional snapshot-based methods, LT-mapper's change management scheme saved approximately 60% of memory and was 8.9 times faster at computing changes between arbitrary sessions.
  • Localization Accuracy: LT-SLAM significantly reduced trajectory drift. In the KAIST dataset, it corrected errors that were present in baseline LiDAR odometry even when no initial alignment between sessions was known.
An example of structural changes over a 1.5-year gap in the KAIST dataset.

Limitations and Future Work

While LT-mapper is a powerful step forward, some areas remain for exploration:

  • Semantic Integration: The current system relies on geometric visibility. Integrating semantic labels (e.g., "this is a tree," "this is a building") could further improve the distinction between seasonal changes and permanent structural ones.
  • Sensor Modality: The framework is currently LiDAR-centric. Extending these "delta map" concepts to visual or multi-modal data could broaden its applicability.
  • Real-time Constraints: While modular and efficient, the full pipeline is designed for "lifelong" updates (often processed offline or in the background) rather than instantaneous real-time map cleaning during high-speed flight or driving.

Summary updated

Abstract

Long-term 3D map management is a fundamental capability required by a robot to reliably navigate in the non-stationary real-world. This paper develops open-source, modular, and readily available LiDAR-based lifelong mapping for urban sites. This is achieved by dividing the problem into successive subproblems: multi-session SLAM (MSS), high/low dynamic change detection, and positive/negative change management. The proposed method leverages MSS and minimizes potential trajectory error; thus, a manual or good initial alignment is not required for change detection. Our change management scheme preserves efficacy in both memory and computation costs, providing automatic object segregation from a large-scale point cloud map. We verify the framework's reliability and applicability even under permanent year-level variation, through extensive real-world experiments with multiple temporal gaps (from day to year).
BibTeX
@inproceedings{kim2022lt,
  author = {Giseop Kim and Ayoung Kim},
  title = {{LT-mapper: A Modular Framework for LiDAR-based Lifelong Mapping}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2022},
  url = {https://openreview.net/forum?id=hx0KmpLGCo},
}

Loading related content...

Figure