Testing different function fitting methods for mobile eye-tracker calibration.

Area of Science:

• Human-Computer Interaction
• Computer Vision
• Biomedical Engineering

Background:

• Eye-tracker calibration establishes a mapping function between extracted features and gaze points.
• Current research focuses on mapping functions, with less attention to parameter estimation methods.
• Noise in mobile eye-tracking, from imprecision or feature detection errors, affects calibration accuracy.

Purpose of the Study:

• To investigate the impact of different fitting methods on eye-tracker calibration accuracy under various noise conditions.
• To compare the performance of commonly used least-squares polynomial regression against alternative methods like ridge regression.
• To identify robust parameter estimation techniques for improved gaze accuracy in real-world scenarios.

Main Methods:

• Developed a binocular gaze simulation incorporating different calibration patterns and noise characteristics.
• Evaluated the accuracy of mapping functions estimated by polynomial regression (least-squares) and ridge regression.
• Quantified performance using mean squared error (MSE) across simulated noise levels.

Main Results:

• Ridge regression consistently outperformed polynomial regression in estimating accurate mapping functions.
• The performance gap widened significantly as simulated data noise increased.
• Outlier-tolerant methods like ridge regression are crucial for noisy calibration data.

Conclusions:

• Ridge regression offers a more robust and accurate alternative to standard polynomial fits for eye-tracker calibration.
• Implementing ridge regression can lead to substantial improvements in gaze estimation accuracy, demonstrated by a 20% reduction in MSE in a mobile eye-tracking experiment.
• Further research into robust fitting methods is essential for advancing mobile eye-tracking technology.