A Novel Method for Scanpath Comparison Based on Levenshtein Distance

Modeling the scanpath of human saccade on presentation pictures is a challenging task. To evaluate the consistence of saccadic eye movement sequences, traditional methods, such as scanMatch, transform scanpath data to string sequence with grid-based method and computer the similarity of string sequences using Needleman-Wunsch algorithm. However, scanpath data are different from string data in some aspects, and grid-based method would cause errors in measurement. In the present work, the difference between two symbol sequences was measure by a modified Levenshtein Distance (LD) method. Here, the pair-point distance was computed from position of different fixation points directly, and the distance was further converted to 0 or 1 by a thresholding method. In this way we can adjust the matching threshold value between points and improved the accuracy of LD method by overcoming measurement error of point around grid boundary. Our experimental results demonstrated that our method had better performance than traditional scanMatch.