Topological Mapping for Mobile Robots Using a Combination of Sonar and Vision Sensing

David Kortenkamp, Terry E. WeymouthAAAI 1994

Abstract

This paper presents a hybrid sensing approach to address the "perceptual aliasing" problem in mobile robot navigation. By fusing low-resolution sonar data for gateway detection with abstracted visual cues (vertical edges) via a Bayesian network, the proposed system enables robots to uniquely identify specific locations, such as doors and intersections, without requiring complex movement sequences. Experimental results demonstrate that while individual sensors may fail to distinguish identical-looking landmarks, their combination achieves 100% recognition accuracy in structured indoor environments.

Introduction

In the early 1990s, mobile robot navigation primarily utilized metric maps or topological maps. While topological maps—represented as graphs of distinctive places—are computationally efficient and more aligned with human spatial reasoning, they are prone to perceptual aliasing. This occurs when different physical locations produce identical sensor signatures (e.g., multiple identical doorways in a hallway). To overcome this, the authors propose a method to identify specific places using only the data available at a single spot by fusing sonar and vision.

Methodology

The proposed architecture centers on the concept of Gateways, which serve as the nodes in a topological network.

Gateway Detection via Sonar

The robot utilizes sonar sensors to follow walls and detect openings such as doors or T-junctions. Upon detecting an opening, the robot centers itself to define a gateway. The authors identified 25 distinct gateway types based on these sonar signatures.

Figure 1: Examples of sonar-defined gateway types, including T-shapes and room entries.

Vision-based Abstracted Scene Representation (ASR)

At each gateway, the robot captures stereo images to extract vertical edges, which are robust features in indoor settings. To account for minor positioning errors, these visual cues are mapped into an Abstracted Scene Representation (ASR)—a 5x5 grid. This abstraction ensures that small variations in the robot's pose do not change the resulting feature set, providing resilience to mechanical inaccuracies.

Figure 2: The Abstracted Scene Representation (ASR) maps visual cues into a grid to handle positioning noise.

Bayesian Fusion

A Bayesian network integrates the likelihoods from both sensors. It combines the probability of a location based on the sonar signature with the likelihood derived from visual cues to calculate the posterior probability of the robot's current node in the map. This allows the system to "overrule" a misclassified sonar reading if visual evidence is strong.

Experimental Evaluation

The system was tested using a Labmate TRC mobile robot in the hallways of the University of Michigan’s Artificial Intelligence Laboratory. The robot traversed three routes five times each to map 16 ASRs across seven unique gateways.

Results and Discussion

The experimental results highlighted the limitations of single-sensor systems:

  • Sonar Alone: Achieved only 57% accuracy, failing to distinguish between three identical "Right Openings."
  • Vision Alone: Correctly identified 4 out of 7 places but struggled with consistency across different runs.
  • Combined System: Achieved 100% accuracy, demonstrating the effectiveness of sensor fusion.

The system also showed robust error recovery; when a gateway was intentionally misclassified to simulate sensor noise, the Bayesian network successfully corrected the localization using visual data.

Conclusion

This research demonstrates that robots do not require high-fidelity computer vision for effective navigation. By focusing processing on "interest points" (gateways) and using abstracted representations, robots can achieve reliable localization in structured environments. Future work will address the challenges of scaling to larger maps and autonomous exploration.

Summary updated

BibTeX
@inproceedings{kortenkamp1994topological,
  author = {David Kortenkamp and Terry E. Weymouth},
  title = {{Topological Mapping for Mobile Robots Using a Combination of Sonar and Vision Sensing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure