Optimizing a dual VAE-classifier method for blink detection in mobile eye-tracking

Precise blink detection is essential for a range of applications, including eye-tracking. Although numerous methods for blink detection have been proposed, their performance may not be optimal under certain complex real-world scenarios. This study introduces a deep learning approach tailored for mobile eye-tracking devices to accurately classify blinks. The dual approach integrates a variational autoencoder (VAE) followed by a classifier that distinguishes between blinks and non-blinks based on the latent vector. To optimize the VAE’s performance, several experimental conditions are explored, including three different data input groupings. Subsequently, to identify the most suitable classifier for the blink detection task, four different classifiers were evaluated. The architecture was tested with two image sizes. The results illustrate that a more complex and diverse dataset requires a higher latent space dimensionality, while image size does not significantly impact performance, indicating that the VAE can capture adequate information even when reduced significantly from the original size. Regarding input data grouping, the additional information from using both eyes or multiple frames does not seem to improve decision-making significantly. Among the classifiers, the K-Nearest Neighbours (KNN) demonstrated the best overall performance. The proposed VAE-classifier method was optimized to achieve an accuracy of more than 99.5%, outperforming other approaches and demonstrating the effectiveness of VAE-based methods for blink detection. Test on clinical videos also demonstrated an excellent agreement with the ground truth. This method surpasses conventional blink detection techniques used in mobile eye trackers, further demonstrating this technique’s potential to support clinical and research eye-tracking applications.