Graph Structure-Based Simultaneous Localization and Mapping Using a Hybrid Method of 2D Laser Scan and Monocular Camera Image in Environments with Laser Scan Ambiguity

Taekjun Oh, Donghwa Lee, Hyungjin Kim, Hyun Myung2015 · DOI

TL;DR

This paper tackles the challenge of robot localization in environments with laser scan ambiguity (like long corridors) by proposing a novel graph-based SLAM approach that fuses data from a 2D laser scanner and a monocular camera. This hybrid method overcomes the limitations of individual sensors, leading to more robust and accurate real-time localization.

What Problem Does This Paper Address?

Robots navigating unknown environments need to know their precise location (localization) and build a map of their surroundings (mapping) simultaneously – a problem known as SLAM. Traditional SLAM methods often rely on laser scanners, which are great for precise distance measurements. However, in environments like long, featureless corridors, laser scans can become ambiguous. The distance to the walls doesn't change much, making it hard for algorithms to distinguish between different locations. This ambiguity can lead to significant errors in pose estimation. While cameras offer rich visual information, they struggle with scale estimation (MonoSLAM) and are sensitive to lighting changes. This paper aims to address the limitations of single-sensor approaches in ambiguous environments by proposing a robust hybrid method.

Methods

The core of the proposed method is a graph-based SLAM framework that integrates data from both a 2D laser scanner and a monocular camera.

  • Hybrid Sensor Fusion: The key innovation lies in fusing depth data from the laser scanner with visual features extracted from the camera. The 3D coordinates of image features are estimated by assuming a flat, vertical wall and using the laser scanner's depth data to augment the 2D image points. This process is made more robust by later rejecting outliers that don't fit the assumed wall geometry.
  • Graph-Based SLAM: The fused sensor data is used to construct a pose graph. Each node in the graph represents a robot pose, and edges represent constraints between these poses derived from odometry, laser scan matching (using GMapping), and the novel hybrid sensor fusion method.
  • Pose Graph Optimization: The final robot pose is estimated by optimizing this graph structure. This optimization process helps to correct accumulated errors from odometry and improve localization accuracy, especially in challenging environments.
  • GMapping Integration: The paper leverages GMapping for its grid-based SLAM capabilities, particularly for generating constraints from laser scan data. However, it acknowledges GMapping's limitations in ambiguous environments and uses its output as one of several constraints within the broader graph optimization framework.

Datasets

The paper describes four experimental datasets collected in various indoor environments:

  • Dataset 1: A long corridor (36.4m x 2.0m).
  • Dataset 2: An L-shaped space (16.0m x 0.9m).
  • Dataset 3: An L-shaped space with parts of the wall made of transparent glass.
  • Dataset 4: Similar to Dataset 2, but with two inclined objects placed to test robustness.

Ground truth positions were acquired through manual measurements.

Key Insights and Ideas

  • Complementary Strengths: Laser scanners and cameras have complementary strengths. Laser scanners provide accurate depth, while cameras offer rich visual features. Combining them effectively can overcome the weaknesses of each.
  • Feature Ambiguity Resolution: The hybrid method's ability to derive 3D feature coordinates using both sensors is crucial for resolving laser scan ambiguity. By using visual features, the system can identify distinct points even when the laser scanner provides uniform depth data.
  • Robustness to Non-Ideal Walls: The paper highlights that the assumption of perfectly flat and vertical walls can be relaxed. The feature matching process effectively rejects outliers, making the system robust to slightly inclined or non-flat surfaces.
  • Graph Optimization for Error Correction: The graph-based SLAM framework, combined with optimization, is essential for correcting accumulated odometry errors and achieving accurate global localization, especially when loop closures are detected.

Key Results

  • Superior Localization Accuracy: The proposed hybrid method significantly outperforms GMapping in terms of localization error, particularly in environments with laser scan ambiguity. RMSE values for the proposed method are consistently smaller across all datasets.
  • Real-time Performance: The algorithm is demonstrated to operate in real-time, with an average update rate of 20 Hz, which is sufficient for practical robot navigation. This is partly achieved by performing computationally intensive image processing only when a new node is added to the graph.
  • Robustness to Ambiguity: Experimental results show that the proposed method can accurately localize the robot in long corridors where GMapping struggles due to uniform laser scan data.
  • Reduced Computational Burden: By accelerating SURF feature extraction with a GPU and processing images only when nodes are added, the algorithm maintains a manageable computational load.

Limitations and Future Work

  • Outdoor Environments: The paper focuses on indoor environments. For future work, the authors propose extending the localization method by applying a probability model for determining 3D feature point coordinates in outdoor environments.
  • Assumption Relaxation: While the method shows robustness to inclined walls, further exploration of how to handle more complex environmental variations might be beneficial.

Summary updated

BibTeX
@misc{oh2015graph,
  author = {Taekjun Oh and Donghwa Lee and Hyungjin Kim and Hyun Myung},
  title = {{Graph Structure-Based Simultaneous Localization and Mapping Using a Hybrid Method of 2D Laser Scan and Monocular Camera Image in Environments with Laser Scan Ambiguity}},
  year = {2015},
  doi = {10.3390/s150715830},
}
Figure