Attention Model Based RNN for Automated Eye-Movement Event Detection
Xiaopan Li, Huazhen Zhang, Shiqian Wu, Hongping Fang · 2021
Eye-movement event detection plays an important role in eye tracking and visual perception understanding. Most of the eye-movement event detection algorithms are mainly implemented by manually setting thresholds or designing features to learn thresholds from raw eye-movement data. In this paper, we propose an attention model based recurrent neural network (AM-RNN) for automated eye-movement event detection. The proposed algorithm can automatically classify the raw eye-movement data into three types (fixations, saccades and post-saccadic oscillations) without manual extraction of features. Specifically, the proposed attention module can automatically assign weight to each feature vector in terms of probability distribution of each feature and thus give perception sights to each feature. Experimental results performed on a large-scale hand mark ground truth dataset (Lund 2013 dataset) show that the proposed method outperforms the state-of-the-art eye-movement event detection algorithms.