Abstract
This paper presents a robust method for the self-calibration of a 2D camera undergoing planar motion by reducing the problem to the calibration of a 1D projective camera. By leveraging the linear properties of the 1D trifocal tensor, the authors circumvent the non-linear instabilities typical of standard 2D self-calibration techniques. Experimental results on both synthetic and real-world data demonstrate high noise tolerance and practical accuracy for robotic applications.
Introduction
Camera calibration is a fundamental task in computer vision, traditionally requiring specialized targets like checkerboards to determine internal parameters such as focal length and principal point. "Self-calibration" aims to recover these parameters from environmental features tracked across multiple images. However, for standard 2D cameras, this often involves solving complex, non-linear equations (like the Kruppa equations) that are highly sensitive to noise.
A common scenario in robotics—planar motion—actually causes these traditional 2D methods to fail due to mathematical degeneracy. This paper addresses this gap by proving that a 2D camera moving in a plane behaves mathematically like a 1D camera, allowing for a more stable, linear solution.
Planar Motion and the 1D Reduction
The core insight of this work is that planar motion (translation within a plane and rotation around an axis perpendicular to it) simplifies the projection geometry. When a camera moves this way, its horizontal coordinates relative to the motion plane can be treated as a 1D camera projection.
By projecting 2D image points onto the "trifocal line"—the intersection of the image plane and the plane of motion—the authors transform a difficult 2D problem into a 1D problem. This reduction is particularly useful for autonomous ground vehicles or robots operating on flat floors, where planar motion is the default state.
The 1D Trifocal Tensor Method
Once the problem is reduced to 1D, the authors utilize the 1D trifocal tensor. Unlike its 2D counterpart, which is algebraically redundant and requires non-linear estimation, the 1D trifocal tensor can be estimated exactly through linear methods.
The calibration process involves:
- Estimating fundamental matrices between views.
- Verifying the planar nature of the motion.
- Projecting points to the trifocal line.
- Solving for the 1D calibration parameters by finding the roots of a cubic polynomial.
This linear approach provides two points on the 2D camera's "Image of the Absolute Conic" (IAC), which is the key to recovering the camera's internal parameters.
Experimental Results
The methodology was validated using both synthetic simulations and real-world robot sequences:
- Synthetic Stability: In tests with a simulated grid, the algorithm showed remarkable resilience. Even with noise levels as high as ±10 pixels, the estimated focal length remained within 0.5% of the true value (401.9 vs. 400).
- Real-World Application: Using a camera on a robot arm, the authors performed a 3D reconstruction of a laboratory scene. The self-calibration method achieved a relative error of only 3.4% compared to a ground-truth reconstruction using a perfect, target-based offline calibration.
Conclusion and Limitations
This paper provides an elegant mathematical shortcut for self-calibration under planar constraints. While highly effective for robotics, the method is limited to planar movement; complex 3D trajectories (like those of a handheld drone) would require different models. Furthermore, at least two distinct planar motions are necessary to recover a full 5-parameter calibration (including skew and aspect ratio). Despite these constraints, the transition from non-linear 2D estimation to linear 1D estimation represents a significant leap in the robustness of self-calibration for mobile robotics.
Summary updated