Non-parametric analysis of eye-tracking data by anomaly detection

A non-parametric approach for distinguishing between individuals by means of recorded eye movements is suggested. The method is based on the principles of stochastic anomaly detection and relies on measured data for probability distribution estimation and evaluation. For visual stimuli that excite the essential nonlinear dynamics of the human oculomotor system, mean gaze trajectories and characterizations of their uncertainty are approximated per individual using eye-tracking data. With this information, eye-tracking profiles are established, against which independently acquired data sets are statistically tested to evaluate the probability that they belong to said profiles. Both Gaussian function fitting and kernel density estimation (KDE) techniques are used for distribution estimation and novel means for testing observations against general distributions are suggested. It is shown that the presented method yields promising results in terms of individual classification based on eye movements. Further, using the KDE method for trajectory distribution estimation provides better selectivity compared to normal distribution fitting.