Abstract
Small mobile robots navigating "urban canyons" often face GPS failure and limited computational resources. This paper proposes a minimalist monocular visual odometry (VO) system that utilizes vertical line pairs as stable landmarks. By modeling error propagation and employing convex optimization—termed Minimum Variance Ego-motion Estimation (MVEE)—the system achieves high accuracy with significantly lower computational overhead than traditional point-tracking methods.
Introduction
In dense urban environments, tall buildings frequently obstruct GPS signals, forcing robots to rely on onboard sensors for localization. Visual Odometry (VO) is a common solution, but standard methods often struggle with the high computational cost of tracking thousands of feature points and the noise introduced by dynamic lighting or moving objects. This research introduces a robust alternative: leveraging the structural regularity of cities by tracking vertical lines.
Methodology: Minimum Variance Ego-motion Estimation (MVEE)
The proposed system focuses on vertical lines—such as building corners and utility poles—which are abundant in urban settings and structurally stable.
The core innovation is the Minimum Variance Ego-motion Estimation (MVEE) method. Unlike simple averaging, MVEE accounts for the reliability of different landmark pairs through three key steps:
- Geometric Modeling: The robot's 2D ground-plane motion is derived from the displacement of vertical line pairs across three consecutive frames using a pinhole camera model.
- Error Awareness: The authors derived a closed-form covariance matrix to predict the expected error for any given pair based on its specific geometry (e.g., proximity of lines).
- Convex Optimization: A quadratic optimization problem assigns higher weights to line pairs with lower predicted variance, ensuring the final estimate is dominated by the most reliable data.
Experimental Evaluation
The method was tested on a custom robot in three distinct urban-style environments. Performance was compared against the Nister point-feature method and the L&L general line-feature method.
Key Results
- Accuracy: MVEE consistently maintained relative trajectory errors under 2%, outperforming both comparison methods across all test sites.
- Efficiency: The system processed steps in 2.3 seconds (MatLab implementation), nearly seven times faster than the point-based Nister method (15.2 seconds). This efficiency is due to tracking roughly 60 robust lines rather than thousands of points.
- Robustness: MVEE achieved an inlier ratio of 42%, far exceeding the 7% seen in point-based tracking, indicating that vertical lines are significantly more reliable landmarks in complex urban scenes.
Conclusion
By simplifying the 3D localization problem into a 2D geometric task constrained by verticality, the authors developed a VO system optimized for small robots with limited processing power. While currently limited to flat-plane motion, the MVEE approach demonstrates that "quality over quantity" in feature selection leads to superior performance in challenging urban environments. Future work aims to extend this to 3D maneuvers by incorporating horizontal features.
Summary updated