Intelligent Eye Movement Perception Method for VDT Visual Fatigue Assessment

Abstract

Video Display Terminal (VDT) visual fatigue has become a significant issue affecting the visual health of the modern population, necessitating a non-invasive and convenient objective monitoring method. This paper proposes an intelligent eye movement perception method based on ordinary cameras, aiming to dynamically monitor key eye movement parameters during VDT use and assess visual fatigue levels. First, a video dataset for blink recognition is constructed, and a Long Short-Term Memory (LSTM)-based blink detection and incomplete blink recognition algorithm is proposed. By extracting the vertical distance from the midpoint of the upper eyelid to the eye corner (DucDuc​) as a temporal feature, high-precision classification of blink events is achieved. Second, to overcome the limitations of manual feature extraction, a hybrid Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) model is designed. This model automatically extracts spatiotemporal features from videos, achieving three-class classification of "non-blink," "complete blink," and "incomplete blink," significantly improving generalization capability and practicality. Third, an image processing-based eyeball center localization and tracking algorithm is proposed. By calculating the dynamic distance change between the eyeball center and the inner eye corner, accurate counting of eyeball rotation frequency is achieved. Finally, based on the blink frequency, complete/incomplete blink count, eyeball rotation frequency, and other parameters extracted using the above methods, combined with subjective questionnaires, the dynamic changes of these parameters during a 120-minute VDT task and their correlation with visual fatigue levels are systematically analyzed. Experimental results show that the proposed LSTM model achieves a blink detection accuracy of 98.44%, the CNN-LSTM hybrid model achieves an average F1 score of 0.93 in cross-subject testing, and the eyeball rotation recognition accuracy is approximately 90%. Correlation analysis reveals that blink count is extremely significantly positively correlated with visual fatigue questionnaire scores (p<0.01p<0.01), while eyeball rotation frequency is extremely significantly negatively correlated (p<0.01p<0.01), validating the effectiveness of the proposed method. This study provides a novel software solution for objective, non-contact VDT visual fatigue assessment with broad application prospects.