Identifying Fixations and Saccades in Virtual Reality

Fixation recognition can significantly reduce the amount of eye movement data to understand human cognitive and visual processing better. Fixation recognition is an essential precondition for eye-based interaction applications in virtual reality. However, the three-dimensional features of virtual reality environments also pose new challenges to existing recognition algorithms that have no consideration of the changes in depth of field. On the other hand, there is a significant imbalance in the distribution of human eye movement behavior, where fixations take up a much more significant proportion of time than saccades. Therefore, when evaluating the effectiveness of recognition algorithms, it is necessary to address the issue of significant differences between positive and negative samples. In response to this issue, this study used a modified F1-score to evaluate the classification performance of various recognition algorithms or models. Based on the modified F1-score, we obtain optimal parameters of three existing threshold-based recognition algorithms, our proposed Velocity, Dispersion, and Vergence-Threshold Identification algorithm, and a deep learning model. We also proposed an outlier process method to reduce the impact of outliers. We compare the performance of these four algorithms with optimal parameters and the deep learning model. The results show that Velocity and Dispersion-Threshold Identification and our proposed algorithm perform the best. The impact of interface complexity on classification results is also preliminarily explored. The results show that the algorithms are not sensitive to interface complexity.