D. Suitt, J. Park, D. Grissom, C. Jones, M. Rickard
California Baptist University,
United States
Keywords: eye tracking, eye gaze, wearables, smart frames
Summary:
We have developed a lightweight, wearable system that supports continuous gaze monitoring in natural, real-world environments. Ocular imagery and environmental data are collected through conventional-looking spectacle frames in traditional form factors (Blink FramesⓇ) that house microcameras, various sensors, supporting microelectronics, and a rechargeable battery. Although the frames have limitations compared to research-grade devices specifically for eye tracking, including lower frame rates and reduced image resolution, they serve as a practical platform for investigating gaze tracking in wearable form factors over time. The potential to adapt a lightweight, consumer-oriented device for reliable gaze estimation presents opportunities for applications in both clinical and real-world settings, where non-obtrusive, extended monitoring may be beneficial. Eye-tracking technology has become an essential tool across medical diagnostics, training, and user experience design, offering insights into eye gaze patterns that enhance applications from clinical assessments to augmented and virtual reality. However, reliable gaze tracking in natural, uncontrolled environments poses challenges, especially when specialized hardware is not viable. By focusing on the Blink Frames platform, our study aims to demonstrate how wearable devices can support eye-tracking research in everyday environments while still maintaining respectable accuracy. Our methodology employs a modular, multi-stage pipeline, incorporating iris localization, eye geometry estimation, and real-world gaze mapping, to capture accurate gaze vectors under varied conditions. Using synthetic data augmentation and a simulated calibration setup, our model is trained to account for challenges such as skew and varied environmental conditions while using relatively limited input data. Our study tracks the iris because it remains reliably visible under varying environmental conditions (unlike the pupil), providing a stable basis for gaze vector estimation without the use of infrared (IR) illuminators. To train our model, we initially annotated approximately 1,000 eye images from diverse environments, subjects, and viewpoints and 14,000 additional images were generated by augmenting the original 1,000 images through transformations; these included random cropping, shifts, scaling, rotations, blurring, noise, brightness adjustment, pixel dropout, occlusion, color shifting, and horizontal flipping. To estimate the gaze direction, we approximated it as a vector originating from the eye center and extending through the iris center. In this model, the eye center acts as a substitute for the fovea, a small, central area in the retina responsible for sharp, detailed vision. We then calibrated the model on real images in which the gaze directions were fixed and then gradually introduced natural eye movement to build a consistent baseline during each test. Our preliminary model performed with an overall mean absolute deviation from ground truth of 0.93 and 0.78 degrees for azimuth and elevation, respectively. Leveraging the Blink Frames system, we have demonstrated a robust gaze tracking methodology that avoids the need for NIR illumination or external reference points.