A Computational Method for Saccade Velocity Profiles Generation

This study presents a computational method for generating eye movement data (specifically horizontal saccade velocity profiles) with the aim of providing benchmark input for the evaluation of data processing workflows and for the employment of machine and deep learning algorithms. The usability of generated saccades is further explored by comparison of two selected algorithms for saccade detection. Precisely, we assess and compare the modified and adaptive I-VT (Identification by Velocity Threshold) algorithm and the MAD (Median Absolute Deviation) algorithm. Both algorithms are tested on the publicly available dataset recorded in one subject while watching a movie trailer and on synthesized saccade velocity profiles. Obtained results suggest that both algorithms have exactly the same performance when generated saccade velocity is not scarcely contaminated with noise (F1Score of 94.42%), while for generated velocity with noise contamination, detection performance of both algorithms significantly degrades (F1 - Score of 36.57% for I-VT and slightly better performance of 40% for MAD algorithm). The proposed method for generating saccadic eye movements profiles could be further utilized to simulate other types of eye movements profiles, such as glissades. Also, the presented algorithm could be explored for the evaluation and design of novel algorithms for eye tracker data analysis, hopefully offering more reliability and higher precision in the eye tracker-based studies.