A Dynamic Threshold-Based Method for Robust and Accurate Blink Detection in Eye-Tracking Data

Blink detection is a critical component of eye-tracking research, particularly in pupillometry, where data loss due to blinks can obscure meaningful insights. Existing methods often rely on fixed thresholds or device-specific noise profiles, which may lead to inaccuracies in detecting blink onsets and offsets, especially in heterogeneous datasets. This study introduces a novel blink detection model that dynamically adapts to varying pupil size distributions, ensuring robustness across different experimental conditions. The proposed method integrates dynamic thresholding, which adjusts based on the mean pupil size of valid samples, Gaussian smoothing, which reduces noise while preserving signal integrity, and adaptive boundary refinement, which refines blink onsets and offsets based on a trends in the smoothed data. Unlike traditional approaches that merge closely spaced blinks, this model treats each blink as an independent event, preserving temporal resolution, which is essential for cognitive and perceptual studies. The model is computationally efficient and adaptable to a wide range of sampling rates, from low-frequency (e.g., 250 Hz) to high-frequency (e.g., 2000 Hz) data, ensuring consistent blink detection across different eye-tracking setups, making it suitable for both real-time and offline eye-tracking applications. Experimental evaluations demonstrate its ability to accurately detect blinks across diverse datasets. By offering a more reliable and generalizable solution, this model advances blink detection methodologies and enhances the quality of eye-tracking data analysis across research domains.