Abstract
In the analysis of point-of-gaze recordings made while observers viewed static stimuli, data are usually made more manageable by identifying fixations and representing the remainder as saccades. Methods for classifying fixations have not been standardized, and often employ velocity, acceleration, duration, or stability thresholds. A problem with these approaches is that there are no commonly accepted values for these thresholds, which may be based on a visual inspection of the data, a subjective process, or on biologically plausible values, which are necessarily imprecise, particularly given variation in the spatiotemporal characteristics of eye-movements across observers and different stimulus types. A simple non-parametric method is proposed here for extracting fixations from point-of-gaze data. The method is velocity-based, but, by contrast with existing methods, the velocity threshold for demarcating saccades was derived automatically from the data for each observer and stimulus individually. Peaks in velocity were obtained from the data producing two populations: one less numerous and with higher velocity peaks, interpreted as saccades, and one more numerous and with lower velocity peaks, interpreted as movements during fixations. The velocity threshold to separate the two populations was then obtained using a method based on the gap statistic, introduced by R. Tibshirani and colleagues as a non-parametric method for identifying the optimum number of clusters in a set of data. To remove some apparently extremely short fixations, thought to represent instrumental noise, a duration threshold was also derived automatically from the data. The whole method was tested on data recorded with a video eye tracker running at 250 Hz while observers viewed static natural scenes for over 32000 one-second trials. Its accuracy in classifying fixations was verified by comparison with fixations labelled by independent experts and against existing parametric techniques.
Supported by EPSRC grant no. EP/F023669/1.