Drowsiness Classification Using ResNet50 and Time Series Transformer Based on Blink Pattern Features

Drowsiness classification faces major technical challenges in accurately capturing long-term temporal patterns while coping with disturbances such as inconsistent lighting and head motion.Traditional approaches based on Eye Aspect Ratio (EAR) analysis or facial landmarks are often susceptible to environmental noise and require complex data preprocessing, which reduces their reliability in real-world conditions.This study proposes a deep learning-based framework that combines ResNet50 and Time Series Transformer (TST) to improve the performance of drowsiness classification.ResNet50 is used to detect eye conditions and generate binary blink patterns, and TST captures the temporal dependencies in these blink patterns.By extracting statistical features such as mean, variance, and blink duration, this approach simplifies the data preprocessing process and improves the model's robustness to environmental disturbances.The experimental results demonstrate that the proposed framework achieves 97% accuracy with comparable precision, recall, and F1 score, outperforming conventional methods in modeling temporal patterns and dealing with technical disturbances.The proposed method exhibits high computational efficiency and provides a practical and reliable solution for real-time drowsiness classification.