Dual-stream temporal–spectral framework for accurate eye movement event detection

Abstract

Accurate detection of eye movement events is crucial for unraveling complex visual behaviors and driving advancements in neuroscience and cognitive research. Deep learning approaches have demonstrated exceptional potential, surpassing traditional and machine learning methods by effectively capturing and modeling intricate patterns in eye movement data. However, existing deep learning methods face limitations, with small receptive fields struggling to capture cross-boundary information and large receptive fields failing to precisely delineate event borders. To address this, we propose GazeFusion, a novel dual-stream framework that integrates global temporal and local spectral features for comprehensive eye movement event detection. The framework comprises a Mamba Temporal Feature Extractor for long-range temporal dependencies, a CNN Spectral Feature Extractor for explicit frequency-domain representation, a Contextualizer leveraging attention mechanisms for global–local temporal–spectral feature fusion, and a Sequential Enhancer for refined sequential modeling. Experimental results on three publicly available datasets consistently demonstrate that GazeFusion outperforms state-of-the-art methods across all categories, offering superior accuracy and robust boundary detection.