Abstract
We address the problem of estimating both translational and angular velocity of a camera from asynchronous point tracks, a formulation relevant to rolling shutter and event cameras. Since the original problem is non-polynomial, we propose a polynomial approximation, classify the resulting minimal problems, and determine their algebraic degrees. Furthermore, we develop minimal solvers for several problems with low degrees and evaluate them on synthetic and real datasets.
Introduction
Modern high-speed sensors like rolling shutter (RS) and event cameras do not capture all pixels simultaneously. RS cameras capture rows sequentially, while event cameras provide a continuous stream of asynchronous brightness changes. When these sensors undergo rapid motion, the resulting data suffers from geometric distortion that standard "global shutter" pinhole models cannot account for. While previous work has addressed absolute pose or motion with known rotation, estimating full 6-Degree-of-Freedom (DoF) motion—unknown translation and unknown rotation—from asynchronous tracks remained a challenging minimal problem.
Methodology
The primary obstacle in solving for 6-DoF motion is the non-polynomial nature of the standard rotation model (Rodrigues formula). To overcome this, the authors utilize Taylor expansions to approximate rotation as a polynomial of degree $K$ (typically $K=1$ or $K=2$).
Problem Classification and Solver Construction
The researchers identified "balanced" minimal problems where the number of constraints matches the number of unknowns, such as 2 observations of 5 points ($m=2, n=5$) or 4 observations of 1 point ($m=4, n=1$). Solvers were implemented using two techniques:
- Gröbner Bases (GB): Optimized for speed, achieving sub-millisecond performance.
- Homotopy Continuation (HC): Used for more complex, higher-degree systems.
A key discovery was "Approximation 2," which uses the transpose of the approximate rotation matrix instead of the inverse. This significantly reduces the number of solutions (e.g., from 120 to 20 in the $m=2, n=5$ case), directly improving RANSAC efficiency.
Experimental Results
The solvers were evaluated on synthetic data, the Carla simulator (providing realistic RS effects), and the real-world Fastec dataset.
Accuracy and Robustness
The proposed solvers demonstrate high numerical stability and maintain accuracy even at high angular velocities. In contrast, the standard Nistér 5-point solver, while effective for pure translation, fails rapidly when rotation is introduced. On the Carla dataset, the $m=2$ solver achieved an Area Under the Curve (AUC) for pose error of 0.74 at a 20° threshold, outperforming the 5-point baseline (0.72).
Computational Efficiency
As shown in the performance table, the Gröbner basis solvers are exceptionally fast, with execution times ranging from 81 to 120 microseconds. This makes them suitable for real-time robotics applications where high-speed motion estimation is critical.
Conclusion and Future Work
This work introduces the first minimal solvers for full 6-DoF motion from asynchronous tracks. While highly effective for small to moderate rotations, the reliance on Taylor expansions means accuracy may degrade at extreme angular velocities unless higher-degree approximations are used. Future research will explore integrating radial distortion parameters and relaxing the constant velocity assumption to handle high-frequency vibrations.
Summary updated