Chronological states of viewer’s intentions using hidden Markov models and features of eye movement

Abstract

To determine the possibility of predicting viewer’s internal states using the hidden Markov model, several features of eye movements were introduced to the model. Performance was measured using the data from a set of eye movement features recorded during recall tests which consisted of observations of three levels of task difficulty. The features were the temporal appearances of fixations and saccades, and combinations of 8 viewed directions during long and short eye movements. As a result, features of long eye movements, such as saccade information, contributed to prediction accuracy. Also, this prediction accuracy was regulated by the difficulty of the task.