Review and Evaluation of Eye Movement Event Detection Algorithms

Area of Science:

• Human-Computer Interaction
• Biomedical Engineering
• Data Science

Background:

• Eye tracking technology analyzes gaze direction, crucial for understanding human behavior and interaction.
• Event detection algorithms classify eye movements like fixations and saccades from raw eye-tracking data.
• Standardized evaluation procedures for comparing these algorithms are currently lacking, hindering progress.

Purpose of the Study:

• To evaluate and compare the performance of various eye-tracking event detection algorithms.
• To assess the impact of threshold values on threshold-based algorithms and determine optimal settings.
• To benchmark machine learning and deep learning algorithms against traditional methods using a unified dataset.

Main Methods:

• Utilized data from a high-speed SMI HiSpeed 1250 eye-tracker system.
• Performed sample-by-sample comparisons of event detection algorithms (threshold-based, machine learning, deep learning) and human coders.
• Focused evaluation on the classification of fixations, saccades, and post-saccadic oscillations.

Main Results:

• All evaluated methods demonstrated good performance in detecting fixations and saccades.
• Significant differences were observed in the classification accuracy among the algorithms.
• Convolutional Neural Networks (CNN) and Random Forest (RF) algorithms generally showed superior performance compared to threshold-based methods.

Conclusions:

• Standardized evaluation using a consistent dataset is crucial for comparing eye-tracking event detection algorithms.
• Machine learning and deep learning approaches, particularly CNN and RF, offer improved accuracy for classifying eye movements.
• Further research into optimized algorithms and evaluation metrics is warranted for advancing eye-tracking analysis.