Eyeblink detection for wearable eye trackers in real-world environments

Although spontaneous eyeblink timing reflects cognitive and psychological states, its potential real-world role as a biomarker remains unclear. The lack of versatile, reliable eyeblink-detection methods for wearable eye trackers constitutes a key challenge. The numerous algorithms that have been developed have inherent limitations. Classic feature-based methods rely on stable imaging conditions, whereas machine-learning-based approaches cannot operate outside the training data distribution. Consequently, videos captured under dynamic lighting conditions, high frame rates, and atypical camera angles fall in the gap between these approaches, which poses analytical challenges. The lack of methods for assessing the prediction quality of unlabeled target data generates concern regarding subsequent analyses. We propose a versatile approach with reasonable accuracy for annotating parts of the target datasets and post-inspections. By leveraging pretrained neural network models, our eye state classifier rapidly improves the prediction quality with increasing amounts of training data and, in subsequent time-series analyses, successfully incorporated the time-series features without further training. We further assessed the detection performance, without labels, by analyzing graphs and estimation statistics to ensure robust predictions. We tested our method on eye videos captured while driving a formula car as an extreme scenario and compared it with an existing algorithm, whereby our approach showed superior accuracy with a modest annotation workload. As the volume of training data increased, our label-free performance metrics improved in tandem with the ground-truth-based metrics; this indicates that the detection quality can be assessed for unlabeled data. Thus, this novel approach supports a range of future eye-tracking research.