Abstract
Gaze tracking is a human-computer interaction technology, and it has been widely studied in the academic and industrial fields. However, constrained by the performance of the specific sensors and algorithms, it has not been popularized for everyone. This paper proposes a single-camera gaze tracking system under natural light to enable its versatility. The iris center and anchor point are the most crucial factors for the accuracy of the system. The accurate iris center is detected by the simple active contour snakuscule, which is initialized by the prior knowledge of eye anatomical dimensions. After that, a novel anchor point is computed by the stable facial landmarks. Next, second-order mapping functions use the eye vectors and the head pose to estimate the points of regard. Finally, the gaze errors are improved by implementing a weight coefficient on the points of regard of the left and right eyes. The feature position of the iris center achieves an accuracy of 98.87% on the GI4E database when the normalized error is lower than 0.05. The accuracy of the gaze tracking method is superior to the-state-of-the-art appearance-based and feature-based methods on the EYEDIAP database.
Introduction
Most high-accuracy gaze trackers currently rely on infrared (IR) light and specialized sensors to create "glints" on the cornea. While effective, these systems are expensive, struggle in bright ambient light (like sunlight), and often fail when users wear glasses due to reflections. Moving gaze tracking to "natural light" using standard webcams is the "holy grail" for making the technology universal. However, this is incredibly difficult; standard cameras often produce low-resolution images where eye features are blurry, and significant head movements can easily break the geometric models used to calculate where a person is looking. This paper introduces a robust gaze-tracking system that works with a standard single camera under natural light, bypassing the need for specialized hardware.
Methodology
The authors propose a regression-based framework that treats gaze estimation as a multi-step geometric problem, divided into a calibration phase and a testing phase.
1. Iris Center Localization (The "Snakuscule")
Instead of simple edge detection, the system uses an area-based circular active contour called a snakuscule. It is initialized using human anatomical priors—specifically the typical ratio of iris size to eye width. The contour then "shrinks or expands" to find the region with the maximum intensity difference between the iris and the sclera (the white of the eye). By using known physical dimensions to initialize the model, the system avoids the "uncontrolled expansion" common in active contour models, making it much faster and more reliable in low resolution.
2. The Stable Anchor Point
Traditional systems use eye corners as a reference (anchor) point. However, eye corners "vibrate" during eye movement and are easily obscured. The authors instead calculate a novel anchor point by averaging the coordinates of 36 stable facial landmarks. This provides a much more consistent reference frame, as moving from a single-point anchor to a 36-point average significantly dampens the "noise" caused by facial expressions or slight tracking errors of individual landmarks.
3. Head Pose Estimation and Mapping
Using six key landmarks (nose tip, chin, mouth corners, and eye corners), the system estimates the head's pitch, yaw, and roll using an iterative 2D-to-3D correspondence algorithm. Second-order polynomial functions then map the "eye vector" (the distance between the iris center and the anchor point) and the head pose to screen coordinates. Finally, a weight coefficient is applied to combine data from the left and right eyes, allowing the system to favor the eye with a clearer line of sight or account for ocular dominance.
Experimental Evaluation
The system was evaluated using two primary benchmarks:
- GI4E Database: Used specifically to test iris center localization, containing 1,236 images from 103 participants.
- EYEDIAP Database: A standard for gaze tracking that includes RGB, RGB-D, and HD video. The authors focused on the "Continuous Screen" and "Discrete Screen" tasks, which involve both static heads and natural head movements.
Results and Discussion
The system achieved 98.87% accuracy on the GI4E database (using a normalized error threshold of 0.05), outperforming previous methods in the same category. On the challenging EYEDIAP dataset, the system achieved a combined gaze error of roughly 6° to 7.6°. While this is higher than IR-based systems (which can be <1°), it represents a significant improvement over previous natural-light, appearance-based models.
In terms of performance, the system runs at 22 fps on the EYEDIAP dataset and up to 35 fps on a standard laptop camera, making it suitable for real-time human-computer interaction. The primary limitation noted is a resolution bottleneck: when the face occupies a small portion of the frame (lower than 512×380), the face detection rate drops significantly. Future work will focus on improving facial landmark tracking in low-quality images and exploring "person-independent" gaze tracking by normalizing features across different eye shapes and sizes.
Summary updated