A Visual-Sensor Model for Mobile Robot Localisation

Matthias Fichtner, Axel GroßmannIJCAI 2003

Abstract

Mobile robot localization remains a fundamental challenge in robotics, particularly when relying on visual data in dynamic environments. This paper introduces a robust visual-sensor model that leverages Hough space representations to match 2D image line segments with a 3D wireframe map. By transforming lines into $(\rho, \theta)$ coordinates, the proposed method achieves significant resilience against lighting fluctuations and partial occlusions. We evaluate two similarity measures—Centred Match Count (CMC) and Grid Length Match (GLM)—demonstrating that the latter provides superior convergence for real-time Monte Carlo Localization (MCL).

Introduction

The "where am I?" problem is central to autonomous navigation. While laser rangefinders and sonar are common, vision-based localization offers a wealth of environmental data. However, traditional vision methods often struggle with high computational costs and extreme sensitivity to lighting conditions. This research addresses these challenges by developing a probabilistic sensor model that is both computationally efficient for real-time applications and robust to the "messiness" of real-world environments, such as truncated lines and unmodeled dynamic objects.

Methodology

The core innovation lies in the transition from the raw image domain to Hough space. The localization pipeline follows a structured sequence of operations to compare observed features with expected map data.

Overview of the visual-sensor model pipeline, illustrating the parallel processing of camera images and 3D world models into a unified similarity measure.
  1. Feature Extraction: Straight-line segments are extracted from camera images after distortion removal.
  2. Hough Representation: Each line is mapped to a point $(\rho, \theta)$ in 2D Hough space. This abstraction ensures that fragmented line segments (e.g., a wall partially obscured by a person) still map to the same region in Hough space.
  3. 3D-to-2D Projection: Using the robot's estimated pose, lines from a 3D wireframe map (VRML format) are projected onto a virtual 2D camera plane.
  4. Similarity Measures: Two primary measures are proposed:
  5. Centred Match Count (CMC): A binary count of image features within a tolerance box of expected features.
  6. Grid Length Match (GLM): A discretized approach that weights features by their total length within Hough space cells, providing a more nuanced comparison.

Experimental Evaluation

The model was tested using a combination of artificial and real-world datasets. Artificial testing involved 15 million random camera poses to analyze the behavior of similarity measures under controlled error. Real-world validation was conducted using a Pioneer 2 robot in a university hallway, collecting over 3,200 images across 64 locations.

The similarity measure distribution plotted against spatial and radial error, demonstrating a sharp, unique peak at the point of zero deviation.

Results indicate that the similarity measure is significantly more sensitive to rotational errors than translational ones. This characteristic is advantageous for wheeled robots, which typically experience higher odometry drift in rotation. The GLM measure consistently outperformed CMC, producing sharper probability distributions and facilitating faster convergence of the localization algorithm.

Conclusion

The proposed Hough-space visual-sensor model offers a robust and efficient solution for robot localization in structured environments. By abstracting geometric features, the system maintains high performance despite environmental noise and lighting changes. While currently optimized for environments dominated by straight lines, such as hallways, future work will explore its application in more organic or cluttered settings and its integration with multi-modal sensor fusion.

Summary updated

BibTeX
@inproceedings{fichtner2003visual,
  author = {Matthias Fichtner and Axel Großmann},
  title = {{A Visual-Sensor Model for Mobile Robot Localisation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2003},
}
Figure