Improving Performance of Eye Movements Classification Using CNN-Transformer Model

In the past decade, eye tracking has been a crucial approach for object selection in digital assistive technology as well as touchless digital signage. Accurate object selection depends on performance of eye movements classification. Many deep learning techniques have been proposed for eye movements classification. Despite of these numerous models, previous approaches have yet to achieve high classficiation accuracy—particularly when dealing with smooth pursuit eye movement. To bridge this scientific gap and improve the effectiveness of eye movement classification, we propose a hybrid CNN-Transformer model. We also incorporated Hyperband hyperparameter tuning to obtain the best parameter values of the model. We evaluated our approach in the GazeCom dataset. This dataset was enhanced with customized annotations designed to accommodate different types of eye movements. Our method yielded F1 scores of 0.9572, 0.9273, and 0.8358 for fixation, saccade, and smooth pursuit eye movements, respectively. The proposed method achieved superior F1 scores by a margin of 1% to 12.36% compared with the state-of-the-art Temporal Convolutional Network (TCN). A significant improvement was observed in the classification of smooth pursuit eye movement. The experimental results imply that the proposed method can serve as a guide for implementing the Transformer models for eye movements classification.