Real-time identification of eye fixations and saccades using radial basis function networks and Markov chains

Analysis of eye-movements is crucial to many applications, from medical diagnosis to gaming. A critical step in this process lies in segmenting raw gaze coordinates provided by the eye-tracker into eye saccade and fixation events. This detection is generally executed offline, as most methods require a complete dataset or large temporal windows. Many of these algorithms also rely on heuristics such as fixed velocity thresholds, yielding variations in the results depending on the user's choice of parameters. To overcome such limitations, we designed a new approach, named RBFNMC, which combines Radial Basis Function Network (RBFN) and Markov Chains (MC) in a Concept Drift framework. Our approach is capable of automatically categorizing saccades and fixations in an online scenario. Comparisons with previous detection techniques revealed accurate predictions, while not requiring fixed threshold parameters. Our results were estimated from real eye-movement datasets collected in experiments with: (i) monkeys in a free-viewing paradigm; (ii) human subjects looking at different types of stimuli. Comparing RBFNMC with several other methods widely cited in the literature, our contribution constitutes a new computational approach to process spatial data streams in an online and unsupervised fashion. As a consequence, we provide an efficient mechanism to detect saccade and fixations which support eye-tracking research in neuroscience and other areas. Finally, it is worth emphasizing that our work additionally provides the possibility of interpreting the decision process by inspecting graph-based visualizations of RBFN and of the transition probabilities in MC. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Analysis of eye-movements plays a crucial role in various applications, but traditional methods have limitations. In this paper, we propose a new approach that can automatically classify eye saccade and fixation events in real-time without the need for fixed threshold parameters. Comparisons with previous methods demonstrate the accurate predictions of our approach. This is significant for eye-tracking research in neuroscience and other fields.