Solution to the SLAM Problem in Low Dynamic Environments Using a Pose Graph and an RGB-D Sensor

Donghwa Lee, Hyun Myung2014 · DOI

TL;DR

This paper tackles the challenge of Simultaneous Localization and Mapping (SLAM) in "low dynamic environments" where objects move infrequently but significantly, causing traditional SLAM to fail. The authors propose a novel RGB-D SLAM method that leverages pose graph optimization, a node grouping strategy, and an error metric to effectively identify and prune false constraints introduced by these object movements, leading to more accurate robot trajectories and maps.

What Problem Does This Paper Address?

Simultaneous Localization and Mapping (SLAM) is a fundamental problem in robotics, enabling a robot to build a map of its environment while simultaneously tracking its own position within that map. Traditionally, SLAM algorithms assumed static environments. However, real-world environments are dynamic, with objects moving. While methods exist for highly dynamic environments (e.g., detecting fast-moving objects), "low dynamic environments" pose a unique challenge. These are situations where objects change positions over longer intervals, making them difficult to detect with standard frame-by-frame object detection. This slow but significant movement leads to "false loop closures" in SLAM, where the robot incorrectly identifies a previously visited location due to the repositioned object. This results in distorted maps and inaccurate robot trajectories. This paper aims to solve the SLAM problem specifically in these low dynamic environments, which are common in many real-world scenarios.

Methods

The proposed method builds upon pose graph SLAM, which represents the robot's trajectory as a graph where nodes are robot poses and edges are constraints between them (e.g., from sensor measurements or dead reckoning). The core contributions are:

  • RGB-D Sensor Integration: The system utilizes an RGB-D sensor (like a Microsoft Kinect) which provides both color and depth information. This is a cost-effective alternative to more expensive laser range finders.
  • Node Grouping: To efficiently handle false constraints arising from object movement, the algorithm groups nodes that observe the same object over time. This grouping is based on covariance merging and specific rules that consider the angle bounds between nodes. This step aims to consolidate information related to potentially moved objects.
  • Error Metric for False Constraint Pruning: An error metric, defined as the average Mahalanobis distance of edges connecting node groups, is introduced. This metric helps identify and quantify the "badness" of constraints between groups. Large error values between groups indicate that the information from one of the groups is likely outdated due to object movement.
  • Pose Graph Re-optimization: Once false constraints (identified by high error metrics between groups) are detected and pruned, the pose graph is re-optimized. This process corrects the robot's trajectory and generates a more accurate map. The algorithm also excludes groups identified as containing outdated information from future considerations.

Datasets

The paper does not mention specific publicly available datasets for training or evaluation. Instead, the authors conducted their own experiments in a laboratory environment. They created "low dynamic environments" by moving tool carts, which served as the objects that changed position. The ground truth for robot pose was obtained using a separate global vision system with a ceiling-mounted camera.

Key Insights and Ideas

  • Low Dynamic Environments are a Distinct Challenge: The paper clearly articulates that slow, infrequent object movements create a unique problem for SLAM that differs from highly dynamic environments. This distinction is crucial for developing targeted solutions.
  • Grouping is Key to Efficiency: Instead of trying to prune individual false constraints, grouping nodes that likely observed the same moved object first allows for more efficient pruning of entire sets of related constraints. This is a significant algorithmic improvement.
  • Covariance Merging for Uncertainty Management: The use of covariance merging to combine uncertainties from different sensor measurements and predictions is a sophisticated way to manage the probabilistic nature of SLAM.
  • Error Metric as a "Movement Detector": The proposed error metric effectively acts as a detector for significant environmental changes by quantifying the inconsistency between different parts of the robot's estimated trajectory.

Key Results

The proposed method demonstrated significant improvements over conventional SLAM in low dynamic environments:

  • Trajectory Accuracy: The median Euclidean distance error relative to the ground truth was significantly reduced. For example, in Experiment 1, the median error for the proposed method was 0.037 m, compared to 0.112 m for conventional SLAM and 0.415 m for odometry alone.
  • Map Quality: The 3D maps generated by the proposed method were more accurate and did not show the distorted or duplicated appearances of objects that were present in the maps from conventional SLAM. The traces of moved objects were effectively removed.
  • Robustness: The method successfully handled situations where objects were moved in different ways (Experiment 1: simple forward movement; Experiment 2: relocation to the side; Experiment 3: multiple objects moved).

Limitations and Future Work

  • 2D Plane Assumption: The current study assumes the robot moves on a 2D plane. The authors suggest that future work should extend the method to handle 6-DOF (six degrees of freedom) movements in 3D spaces.
  • Sensor Errors: The method assumes that erroneous constraints generated by sensor noise have been removed by robust sensing techniques prior to the pose graph optimization. It does not explicitly handle sensor failures or extreme noise.
  • Limited Landmark Movement: The grouping method assumes that inter-group edges are not expected to be large unless landmarks move. This might be a limitation if landmarks themselves are highly dynamic or if there are other sources of large inter-group errors not related to object movement.

Summary updated

BibTeX
@misc{lee2014solution,
  author = {Donghwa Lee and Hyun Myung},
  title = {{Solution to the SLAM Problem in Low Dynamic Environments Using a Pose Graph and an RGB-D Sensor}},
  year = {2014},
  doi = {10.3390/s140712467},
}
Figure