A Life-Long SLAM Approach Using Adaptable Local Maps Based on Rasterized LIDAR Images

Waqas Ali, Peilin Liu, Rendong Ying, Zheng Gong

TL;DR

This paper introduces a lightweight, "life-long" SLAM (Simultaneous Localization and Mapping) system that represents 3D LIDAR data as a collection of 2D rasterized images. By treating local maps as images, the authors leverage efficient computer vision techniques like "Bag of Visual Words" (BoW) to achieve high-accuracy loop closure with significantly lower CPU and memory overhead than current state-of-the-art methods.

What Problem Does This Paper Address?

Standard SLAM algorithms face a "memory explosion" problem: as a robot operates over long periods or multiple sessions in the same area, the map grows indefinitely. This leads to massive computational requirements for tracking and loop closure (recognizing a previously visited location).

Furthermore, real-world environments are dynamic—furniture moves, people walk by, and seasons change. A "life-long" SLAM system must be able to update its map with new information while gracefully "forgetting" or culling redundant and outdated observations to keep the system responsive and efficient.

Methods

The authors propose a system architecture that bridges the gap between 3D LIDAR point clouds and efficient 2D image processing:

  1. Rasterization: Instead of storing raw 3D points, the system projects LIDAR data into 2D grayscale images. The pixel intensity represents the elevation ($z$-value), effectively creating a "top-down" visual summary of a local area.
  2. Local Map Representation: The global map is managed as a graph of these rasterized local maps rather than one giant point cloud.
  3. Bag of Visual Words (BoW) for LIDAR: Typically used in visual SLAM (cameras), the authors apply BoW to their rasterized LIDAR images. They extract ORB features from these images to build a searchable database. This allows the robot to quickly "query" the database to see if its current view matches a previous local map (loop closure).
  4. Map Management & Culling: To prevent the map from growing forever, the system includes a culling algorithm. It removes redundant local maps based on distance traveled and updates existing ones when the robot revisits an area, ensuring the "online" map remains lightweight.

Datasets

The researchers evaluated their approach using two primary benchmarks:

  • KITTI Odometry Dataset: A standard outdoor benchmark containing 22 sequences of 3D LIDAR data from a car driving through urban and rural environments.
  • Indoor Lab Dataset: A custom dataset recorded with a VLP-16 scanner in a dynamic lab environment over multiple runs to test long-term performance and map stability.

Key Insights and Ideas

  • The 2D Shortcut: The most clever insight is that for many navigation tasks, a full 3D point cloud is overkill. By rasterizing the data, the authors can use highly optimized image-matching algorithms (like DBOW2) that are orders of magnitude faster than 3D point cloud matching.
  • Efficiency through Sparsity: By only using a limited number of local map images to build the BoW database—rather than every single keyframe—the search space for loop closure is dramatically reduced without sacrificing much accuracy.
  • Decoupling Online and Offline: The system maintains an "online" lightweight graph for immediate navigation while saving marginalized data to an "offline" database, striking a balance between immediate performance and long-term memory.

Key Results

  • High Accuracy: The loop closure system achieved recall and precision rates above 90% on the KITTI dataset.
  • Low Computational Footprint: In head-to-head comparisons with popular frameworks like ORB-SLAM2 and LOAM, this method showed a massive reduction in resource usage:
  • CPU Usage: Remained consistently under 10%, whereas competitors often spiked to 40-50%.
  • Memory Usage: Stayed below 5-8%, while some state-of-the-art systems (like ISCLOAM) consumed up to 60-80% of available memory as the run progressed.
  • Localization: The relative translation and rotation errors were on par with or better than LOAM and VINS, proving that the efficiency gains didn't come at the cost of basic navigation accuracy.

Limitations and Future Work

While the paper demonstrates excellent results, a few areas remain for exploration:

  • Vertical Complexity: Rasterization into a single 2D plane might lose critical information in complex multi-story environments or areas with significant overhanging structures (like multi-level parking garages).
  • Feature Dependency: The system relies on ORB features extracted from the rasterized images. In very sparse or featureless environments (like a long, smooth hallway), the "visual" words might not be distinct enough for reliable loop closure.
  • Semantic Integration: The authors mention that future work could involve adding semantic labels (e.g., "car," "tree") to the rasterized images to further improve the robustness of the map management system.

Summary updated

Abstract

Most real-time autonomous robot applications require a robot to traverse through a dynamic space for a long time. In some cases, a robot needs to work in the same environment. Such applications give rise to the problem of a life-long SLAM system. Life-long SLAM presents two main challenges i.e. the tracking should not fail in a dynamic environment and the need for a robust and efficient mapping strategy. The system should update maps with new information; while also keeping track of older observations. But, mapping for a long time can require higher computational requirements. In this paper, we propose a solution to the problem of life-long SLAM. We represent the global map as a set of rasterized images of local maps along with a map management system responsible for updating local maps and keeping track of older values. We also present an efficient approach of using the bag of visual words method for loop closure detection and relocalization. We evaluate the performance of our system on the KITTI dataset and an indoor dataset. Our loop closure system reported recall and precision of above 90 percent. The computational cost of our system is much lower as compared to state-of-the-art methods. Our method reports lower computational requirements even for long-term operation.
BibTeX
@article{ali2021life,
  author = {Waqas Ali and Peilin Liu and Rendong Ying and Zheng Gong},
  title = {{A Life-Long SLAM Approach Using Adaptable Local Maps Based on Rasterized LIDAR Images}},
  journal = {IEEE Sensors Journal},
  year = {2021},
  doi = {10.1109/jsen.2021.3100882},
}

Loading related content...

Figure