Using the Mean Shift Algorithm to Make Post Hoc Improvements to the Accuracy of Eye Tracking Data Based on Probable Fixation Locations
Yunfeng Zhang
Committee: Anthony Hornof (chair), Ed Awh, Arthur Farley, Michal Young
Directed Research Project(Jun 2010)
Keywords: eye tracking; error correction

If they choose to look for it, eye tracking researchers will almost always see disparities between the participantsâ actual gaze locations and the locations recorded by the eye trackers. Sometimes these discrepancies are so great that they dramatically affect the validity of the theoretical and empirical claims made based on the eye tracking data. Much of the disparity is in fact a type of eye tracking error—systematic error—which tends to stay constant over time. A challenge in identifying the size and direction of the systematic error is to determine the participants' actual gaze locations from the raw data. Mapping gazes to incorrect locations (not their actual locations) would result in misleading disparities and hence inaccurate estimate of the systematic error. In this paper, we propose a general method that can reliably reduce the systematic error and restore the eye movements to their true locations. The method addresses the difficulty in finding mappings between gazes and their correct locations by embracing a typical characteristic of the eye movement data—that the disparities of the correct mappings tend to be similar to each other and hence they form the highest density cluster among all disparities. The method then uses a variant of the mean shift algorithm to locate the cluster and its center, and to reduce the errors by subtracting the center disparity from the eye movement data. This paper presents the method, an extended demonstration, and a validation of the efficacy of the error correction technique.