EllSeg: An Ellipse Segmentation Framework for Robust Gaze Tracking.

Area of Science:

• Computer Vision
• Biomedical Engineering
• Machine Learning

Background:

• Ellipse fitting is crucial for pupil and iris tracking in video oculography.
• Traditional methods struggle with occlusions from eyelids, eyelashes, or camera angles, breaking ellipse fitting algorithms.
• Existing eye part segmentation techniques often yield incomplete edge data, hindering accurate ellipse fitting.

Purpose of the Study:

• To develop a robust method for direct segmentation of elliptical eye structures (pupil and iris).
• To evaluate the performance of a convolutional neural network (CNN) based direct segmentation approach.
• To compare the proposed method against standard eye part segmentation for pupil and iris tracking.

Main Methods:

• Training a convolutional neural network (CNN) to directly segment entire elliptical pupil and iris structures.
• Utilizing previously segmented eye parts generated by various computer vision techniques as input.
• Testing the framework on publicly available synthetic segmentation datasets.

Main Results:

• The CNN-based direct segmentation framework demonstrates robustness to occlusions.
• Achieved at least a 10% increase in pupil center detection rate within a two-pixel error margin.
• Achieved at least a 24% increase in iris center detection rate within a two-pixel error margin.
• Outperformed standard eye part segmentation methods in tracking accuracy.

Conclusions:

• Directly segmenting elliptical structures with a CNN offers superior performance for eye tracking.
• The proposed method effectively handles occlusions, a common challenge in video oculography.
• This CNN-based approach enhances the reliability and accuracy of pupil and iris tracking systems.