Abstract
Estimating the relative motion of a vehicle is a fundamental task in autonomous navigation. Traditional "Structure from Motion" techniques often require multiple point correspondences and computationally expensive RANSAC loops to filter outliers. This paper introduces a novel approach that leverages the planar motion constraints of ground vehicles to estimate relative pose using only a single affine correspondence. By utilizing the local shape information inherent in affine matches, the proposed solvers (1AC and 1ACf) achieve real-time performance and high robustness. Experimental results on synthetic and real-world datasets demonstrate that this method significantly outperforms traditional multi-point solvers in both speed and accuracy.
Introduction
In computer vision, determining camera motion between frames typically involves solving for the "essential matrix" using at least five matching points. However, for autonomous vehicles, this general approach is often overkill and inefficient. Vehicles primarily move on a 2D ground plane, which reduces the degrees of freedom (DoF) from five to just two: rotation and translation direction.
Furthermore, real-world urban environments are challenging due to high outlier ratios caused by moving objects and repetitive textures. The more matches an algorithm requires to generate a motion hypothesis, the higher the likelihood of selecting an outlier, leading to longer processing times in iterative frameworks like RANSAC. This research addresses these challenges by developing a "minimal solver" that exploits planar constraints and high-density information from affine correspondences.
Methodology
Affine Correspondences and Planar Constraints
The core innovation lies in the use of Affine Correspondences (ACs) instead of simple point pairs. While a standard point match provides only the target coordinates, an AC describes how the local neighborhood around a pixel is warped (scaled, rotated, or sheared). This additional information provides enough linear constraints to solve the reduced-DoF planar motion problem using just one match.
The authors introduce two specific solvers:
- 1AC: Designed for fully calibrated cameras, solving for rotation and translation from one AC.
- 1ACf: Designed for semi-calibrated cameras where the focal length is unknown, recovering both motion and focal length from a single AC.
Weighted Histogram Voting
To replace the traditional RANSAC loop, the authors propose a Weighted Histogram Voting scheme. Since only one match is needed to generate a motion hypothesis, every match in the image can "vote" for a specific motion path in a parameter space. The peak of the resulting histogram represents the most likely true motion. This approach allows for outlier rejection in a single pass, making the computational cost linear relative to the number of points.
Experimental Evaluation
The methods were evaluated using both synthetic environments and the Malaga Urban Dataset, which contains several kilometers of urban driving video with GPS ground truth. The synthetic tests allowed the researchers to measure resilience to "planar noise"—deviations from a perfectly flat road—while the Malaga dataset provided a benchmark for real-world performance.
Key Results
The proposed 1AC solver demonstrated a massive leap in efficiency. On the Malaga dataset, the 1AC + Histogram method processed frames in approximately 16ms, compared to 588ms to 749ms for standard 3-point or 5-point RANSAC-based solvers.
Beyond speed, the 1AC solver consistently produced lower angular errors for rotation and translation. Even when the planar constraint was slightly violated (up to 1 degree of tilt), the 1AC solver remained more accurate than general-purpose solvers. Additionally, the 1ACf solver proved numerically stable in recovering camera focal lengths, a task that is notoriously difficult for minimal solvers under noisy conditions.
Discussion and Conclusion
By reducing the required number of matches to one, this research significantly improves the robustness and speed of visual odometry for ground vehicles. The primary limitation remains the planar assumption; the algorithm is specifically tuned for ground-based platforms and would require modification for aerial or handheld use cases where motion is not constrained to a plane. Future work may involve integrating deep-learning-based affine feature extractors to further enhance the front-end of the pipeline, potentially making the system even more resilient to complex urban textures.
Summary updated