Accurate Model-Based Point of Gaze Estimation on Mobile Devices

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Biomedical Engineering

Background:

• Remote Point of Gaze (PoG) estimation typically uses infrared cameras and models, assuming a constant relative eye-to-system roll (R-Roll).
• This assumption is violated in mobile eye-tracking systems, where significant R-Roll variations occur.
• Existing models developed for desktop systems exhibit reduced accuracy on mobile devices due to unaddressed R-Roll effects.

Purpose of the Study:

• To analyze the impact of R-Roll angle and viewer's visual-optical axis offset on PoG estimation accuracy for mobile devices.
• To develop and validate a novel PoG estimation method that compensates for R-Roll variations.
• To enhance the performance of eye-tracking systems on hand-held mobile devices.

Main Methods:

• Analysis of PoG estimation accuracy dependence on R-Roll angle and angular offset.
• Development of a new PoG determination method incorporating R-Roll compensation.
• Experimental validation using a prototype infrared smartphone eye-tracking system.

Main Results:

• The new method achieved approximately 1° accuracy at a 90° R-Roll angle.
• Conventional methods without R-Roll compensation achieved 3.5° accuracy under similar conditions.
• Experimental errors correlated with R-Roll angle magnitude, validating the analysis.

Conclusions:

• R-Roll angle significantly affects PoG estimation accuracy in mobile eye-tracking.
• The proposed R-Roll compensating method substantially improves PoG estimation accuracy.
• This advancement offers significant potential for more reliable mobile eye-tracking applications.