An algorithm for determining clusters, pairs or singletons in eye-movement scan-path records

Abstract

Progress in the development of techniques and equipment to record eye movements (Young & Sheena, 1975) has not been matched by progress or innovation in analytic techniques for assessing eye-movement data itself. By and large, studies that employdynamic eye-movement r~~ords as a p~ysiological measure of attentionor of cogrutive processing focus on fixations or saccade amplitudes as key variables. Yet the more global feature of eyemovement records, the overall distribution of fixations on the visual two-space (the actual stimulus display) has not received as muchattention. An appeal to topographic, g~ap~-t~eore~ic, or spatial-point type of analysis can proVIde insights mto the distributional structure of such global dimensions of eye-movement data.A visual scene can be regarded as a two-dimensional space that is characterized by regions that differ along a nurr.tber of dimensions: informative value, texture, luminosity, and so forth. We can ask whether the successive displacement of eye positions, reflecting the global scan patterns of subjects, is regular, random, or structured in patterns of clusters, pairs, or single fixations over the visual field. Givena technique for reliably answering such questions for large data sets, we can then develop a number of indexes to characterize the ocular behavior of subjec~s in waysthat go beyond the analysis of typically static variables, suchas fixation durationand saccade amplitude.We present here a simplealgorithm that has varied applications in the analysis of eye-movement scan-path records. This algorithmis based in part on techniques developed for use in ethological and taxonomic studies of behavior. The algorithm describedhere departsfrom such techniqu~s in one important respect. Althoughclustering sc.hemes m the past have not been particularly concerned WIth the order of arrival of elements in a cluster, this is The program has three phases: initialization, execution, and output. During initialization, arrays for clusters, pairs, and singletons are dimensioned, and each array cell is given a starting value of zero. Each array stores the total duration of fixations, and the cluster and pair arrays also hold values for mean duration and mean x,y location for e~ch pair or cluster. Furthermore, the cluster array contams th~total number of fixations forming any given c.luster in the array. The user provides answers to questions that set values to two variables. These variables are the minimum number of fixations needed to establish a cluster and the maximum distancepermitted between fixations to determine whether fixations are close enough to form a cluster or to be included in an ongoing cluster. User-defined values for these variables allow for flexibility in the cluster definitionto fit the nature of data and the questions asked in an analysis of that data. The values of these variables depend in large measure on the nature of the visual stimulus field and the size and densityof that field.As shown in Figure 1, the execution phase begins with program CLSTDR's reading ...

Visual Exploration of Eye Movement Data Using the Space-Time-Cube

“…Typically a time stamp (temporal data) and gaze point location within the configured screen coordinate system (spatial data) is reported by the tracker. One of the challenges, as tackled by many researchers [2,6,7,8,9,10] is how to process, manage, and use these continuous streams of data efficiently and effectively to support a usability evaluation.…”

Recording and analyzing eye-position data using a microcomputer workstation

“…Thus, the exact spatial and temporal sequence in which the eye samples the image during a trial is a crucial factor that is incorporated into our fixation and clustering algorithms. Our method of defining fixations and clusters by taking into account the scan path is uniquely different from most alternative clustering approaches in which fixation data are grouped on the basis of the density of eye-position data at various spatial locations on the image without regard to scan-path information (see, e.g., Latimer, 1988, andScinto &Barnette, 1986, for different approaches).…”

Eye-gaze-contingent control of the computer interface: Methodology and example for zoom detection

“…As opposed to the more subjective techniques often used for cluster eyemovement data (e.g., Belofsky & Lyon, 1988;Ramakrishna et al, 1993;Scinto & Barnette, 1986), the MST-based technique allows automated clustering based upon controlled comparison of local samples of eye movements. Efficient clustering algorithms (e.g., Camerini et al, 1988) have already been developed in other application areas, easing some of the burdens of developing new methodologies.…”

“…Cluster characterization variables include height and width, mean fixation position (both unweighted and weighted by time), area, and other indices. Scinto and Barnette (1986) used a similar subjective strategy for deciding whether fixations were part of the same cluster. They specified the minimum number of fixations required to establish a cluster and the minimum distance permitted between fixations before separation into multiple clusters.…”