An eye movement analysis algorithm for a multielement target tracking task: Maximum transition-based agglomerative hierarchical clustering

An algorithm was developed to characterize, compare, and analyze eye movement sequences that occur during visual tracking of multiple moving targets. When individuals perform a task requiring interrogating multiple moving targets, complex and long eye movement sequences occur, making sequence comparisons difficult in whole and in part. The developed algorithm characterizes a sequence by hierarchically clustering the targets that an individual interrogated through an unordered transition matrix created from the frequencies of eye fixation transitions among the targets. Then, the resulting sets of clustered targets, which we define as multilevel visual groupings (VGs), can be compared with analyze performance. The algorithm was applied to an aircraft conflict detection task. Eye movement data were collected from 25 expert air traffic controllers and 40 novices. The task was to detect air traffic conflicts for easy, moderate, and hard difficulty scenarios on simulated radar display. Experts' and novices' multilevel (level one composed of pairs, and level two composed of three or four targets) VGs were aggregated and visualized. Chi-square tests confirmed that there were significant differences for easy (level one: p < 0.001, level two: p = 0.004), moderate (level two: p = 0.047), and hard (level two: p < 0.001) difficulty scenarios. The algorithm supported identifying different eye movement characteristics between experts and novices. Scans of the experts had multilevel VGs around the conflict pairs, whereas those of the novices included different aircraft. The results show promise for using the compact representation of eye movements for performance analysis.