Minimal sequential gaze models for inferring walkers' tasks

Eye movements in extended sequential behavior are known to reflect task demands much more than low-level feature saliency. However, the more naturalistic the task is the more difficult it becomes to establish what cognitive processes a particular task elicits moment by moment. Here we ask the question, which sequential model is required to capture gaze sequences so that the ongoing task can be inferred reliably. Specifically, we consider eye movements of human subjects navigating a walkway while avoiding obstacles and approaching targets in a virtual environment. We show that Hidden-Markov Models, which have been used extensively in modeling human sequential behavior, can be augmented with few state variables describing the egocentric position of subjects relative to objects in the environment to dramatically increase successful classification of the ongoing task and to generate gaze sequences, that are very close to those observed in human subjects.