Eye gaze direction is a powerful cue for users' intent. However, it is difficult to interpret in natural situations, since gaze serves multiple purposes. Here, we demonstrate that by modeling different gaze behaviors and the transitions between them during a cursor guidance task that includes an obstacle avoidance constraint using a Hidden Markov Model, we can infer the users' goal out of a field of 49 possibilities. Users are not given any specific instructions regarding their gaze, and typically spend only a small fraction of the time looking at their intended target. Nonetheless, our experimental results indicate that the hidden Markov model for gaze enables reliable user independent identification of the target of the cursor movement. The accuracy with which the target region...[ Read more ]

Eye gaze direction is a powerful cue for users' intent. However, it is difficult to interpret in natural situations, since gaze serves multiple purposes. Here, we demonstrate that by modeling different gaze behaviors and the transitions between them during a cursor guidance task that includes an obstacle avoidance constraint using a Hidden Markov Model, we can infer the users' goal out of a field of 49 possibilities. Users are not given any specific instructions regarding their gaze, and typically spend only a small fraction of the time looking at their intended target. Nonetheless, our experimental results indicate that the hidden Markov model for gaze enables reliable user independent identification of the target of the cursor movement. The accuracy with which the target region is identified increases over time, eventually surpassing 80%. We applied this model to a human machine interface for a robotic arm reaching task. We show that performance in target identification reaches 80%. In comparison with a rule based algorithm, which makes use of less history information, the HMM based inference improves performance significantly, suggesting that proper modeling of the entire gaze trajectory is critical in analyzing gaze.

Keywords - Gaze, Eye Tracking, Hidden Markov model, Intent