Analysis of Visual Sequences for Driver Drowsiness Detection

Driver drowsiness remains one of the leading causes of traffic accidents, highlighting the urgent need for accurate and non-invasive monitoring systems. This study proposes a hybrid deep learning framework that integrates Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks to detect drowsiness from facial visual sequences. Four prominent CNN architectures, ResNet50, ResNet101, InceptionV3, and DenseNet121, are utilized to extract spatial features from video frames, while the LSTM component captures temporal dependencies to model the progression of fatigue-related facial dynamics. The evaluation leverages the NTHU Driver Drowsiness Detection (NTHU-DDD) dataset, which encompasses challenging visual conditions such as varying illumination and the use of glasses or sunglasses. Preprocessing steps include facial landmark detection, cropping, normalization, and construction of sequential input. Among all tested models, DenseNet121 demonstrates the most robust performance, achieving 94.63% accuracy and exhibiting high recall in detecting drowsy states. The findings underscore the model’s strong generalization capabilities across diverse real-world conditions, positioning DenseNet121 as a highly promising candidate for deployment in vision-based Advanced Driver Monitoring Systems (ADMS).