Improving gaze accuracy and predicting fixation in real time with video based eye trackers

摘要

Studies of eye movements require accurate gaze, fixation and saccade detection, and most recent studies use video based eye-trackers for this purpose. We present two methods which significantly improve current eye tracking technology, with only minor additions to standard experimental protocols. First, for video-based eye trackers, we characterize a significant pupil-size dependent artifact which systematically biases reported gaze position. By varying display luminance while subjects maintain fixation, we observe corresponding changes in pupil size inducing a gaze position error and obtain an empirical solution to correct it. Applying our technique in software to a commercial video-based eye tracker, we obtain a substantial improvement in the accuracy of gaze position. After correction, the standard deviation of gaze positions around a point of fixation during a 10 second interval reduces by as much as 7.5?? and 5.9?? in the worst case, with an average reduction of 2.29?? and 2.95?? across subjects (n = 6) and screen positions (m = 9), for horizontal and vertical directions, respectively. Additionally, we describe a simple yet effective method for predicting the next fixation during a saccade in flight. Leveraging the relationship between peak velocity and time left in a saccade, we are able to fit model parameters to individual subjects and then use on-line velocity data to predict future fixations. To evaluate the scheme, subjects free-viewed a four minute introduction of a nature documentary. For a stimulus display refresh rate of 100 Hz, we correctly predict fixation onsets to within a frame 95% of the time. Our methodology improves gaze accuracy and allows experimenters a direct access to a window of time immediately around the onset of fixation, opening the door for gaze and saccade contingent experiments using current commercial eye trackers.