Introduction
Visual odometry is a cornerstone of autonomous navigation, typically requiring the matching of multiple point features across frames to estimate camera motion. However, ground vehicles operate under non-holonomic constraints—they cannot move sideways. By exploiting these constraints through the Ackermann steering model, which assumes a vehicle moves along a circular arc, motion can theoretically be solved using just a single feature across two views.
While elegant, standard two-view, one-point solvers are notoriously sensitive to image noise. They also struggle to incorporate line features, which are prevalent in urban environments. This paper introduces the 1FNV (1-Feature n-View) solver, a robust method that tracks a single feature across multiple frames to significantly improve motion estimation accuracy.
Methodology
The core of the 1FNV solver lies in extending the traditional Ackermann model over a temporal window.
Extended Ackermann Model
The authors assume that over a short sequence of $n$ frames, the vehicle maintains a locally constant velocity and steering angle. This implies that the relative transformation between any two consecutive frames is identical.
The n-View Constraint
By stacking observations of a single feature (either a point or a vertical line) across $n \ge 3$ views, the authors derive an $n$-linear constraint. For vertical lines, the problem is simplified by projecting them into the horizontal plane, allowing them to be treated mathematically as points.
To solve for the rotation angle $\theta$, the complex rank-minimization problem is reduced to finding the roots of a univariate polynomial. This is achieved using Viète’s formulas and the Weierstrass substitution. A histogram voting scheme then identifies the best consensus among multiple feature candidates, followed by a 1D line search refinement.
Experimental Results
The 1FNV solver was evaluated using both synthetic simulations and the real-world KITTI dataset.
Synthetic Analysis
Simulations demonstrated that the 1FNV method's robustness increases dramatically with the number of views. While a 2-view solver is highly susceptible to noise, the 1FNV solver acts as a temporal filter, improving the signal-to-noise ratio as more frames are integrated.
Real-World Performance
On the KITTI dataset, the 1FNV method outperformed the 1-point baseline. For instance, on sequence 0046, the 1FNV point-based solver achieved a rotation RMSE of 0.2735, compared to 0.3338 for the baseline. The line-based variant was particularly successful in feature-poor urban areas; on sequence 0095, it yielded a median translation error of 0.0427, nearly doubling the accuracy of the point-based baseline (0.0890).
Conclusion
The 1FNV solver demonstrates that tracking a single feature over multiple views provides a powerful and efficient way to estimate the motion of non-holonomic vehicles. By maintaining a univariate formulation, the method remains computationally efficient while gaining substantial robustness against noise. Future work aims to relax the constant velocity assumption by modeling linear changes in orientation over time.
Summary updated