Abstract: Video-based glint-free eye tracking commonly estimates gaze direction based on the pupil center. The boundary of the pupil is fitted with an ellipse and the Euclidean center of the ellipse in the image is taken as the center of the pupil. However, the center of the pupil is generally not mapped to the center of the ellipse by the projective camera transformation. This error resulting from using a point that is not the true center of the pupil directly affects eye tracking accuracy. We investigate the underlying geometric problem of determining the center of a circular object based on its projective image. The main idea is to exploit two concentric circles—in the application scenario these are the pupil and the iris. We show that it is possible to compute the center and the ratio of the radii from the mapped concentric circles with a direct method that is fast and robust in practice. We evaluate our method on synthetically generated data and find that it improves systematically over using the center of the fitted ellipse. Apart from applications of eye tracking, we estimate that our approach will be useful in other tracking applications.
Introduction
In video-based eye tracking, researchers usually estimate gaze by finding the center of the pupil. Standard practice involves fitting an ellipse to the pupil's boundary and using the geometric center of that ellipse as the pupil's location. However, there is a fundamental geometric "trap": under a perspective transformation (how a camera sees the world), the center of a circle does not map to the geometric center of the resulting ellipse. This discrepancy introduces systematic errors in gaze tracking, especially when the eye is viewed at an angle. While some systems try to correct this using complex 3D models or iterative optimization, these are often computationally expensive or require specialized hardware.
Methods: A Direct Geometric Solution
The proposed approach leverages the geometric relationship between two concentric circles—the pupil and the iris—to recover the true center. Unlike previous methods that might require iterative search or multiple camera views, this approach is a "one-shot" calculation that is both fast and mathematically robust.
The Eigenvalue Solution
The authors represent the two detected ellipses (pupil and iris) as $3 \times 3$ symmetric matrices, $Q_1$ and $Q_2$. In projective geometry, the set of all linear combinations of these two matrices is called a "pencil of conics." The problem of finding the true center reduces to an eigenvalue problem for the matrix $A = Q_2 Q_1^{-1}$:
- Eigenvectors: The eigenvectors of this matrix reveal the coordinates of the true projective center $p$.
- Eigenvalues: The eigenvalues allow for the calculation of the ratio of the radii ($R/r$) of the original circles.
This ratio is a projective invariant that can be recovered even if the camera's exact position and focal length are unknown.
Experimental Validation
The study used a established 3D eye model to render realistic eye images, allowing for precise control over camera rotation (polar and azimuthal angles) and pupil size. This synthetic data provided a ground truth to compare the estimated center against the standard Euclidean center.
Key Results
- High Accuracy: In synthetic tests, the proposed method consistently achieved an estimation error of less than one pixel, significantly outperforming the standard "Euclidean center of the fitted ellipse" approach.
- Robustness to Angle: As the camera angle increases (creating more perspective distortion), the error of standard ellipse-fitting grows, while the concentric circle method remains stable and accurate.
- Pupil Size Sensitivity: While all methods struggle slightly more with very large pupils due to increased distortion, the proposed method remained the most accurate across all tested sizes.
Limitations and Future Work
The primary limitation of the current model is the assumption of a refractive index of 1.0. In real eyes, light is bent by the cornea (refraction), which further distorts the pupil's appearance. Additionally, in real-world scenarios, the iris is often partially covered by eyelids, making fitting a perfect iris ellipse difficult.
Future work suggests that because the iris is less affected by refraction than the pupil, the recovered "radii ratio" could potentially be used to implicitly model and correct for corneal refraction in future iterations of the algorithm.
Summary updated