A Novel Approach to Detect Driver Drowsiness Using Transfer Learning and Hybrid Features

Abstract

Abstract In worldwide, drowsiness is one of the prevalent reasons to cause accident. Statistics show that fatigued drivers are a major factor in causing many accidents. According to studies by the National Sleep Foundation, 20% of drivers feel sleepy to some extent while driving. Deep learning-based methods are the most recent ones that researchers have used to analyse videos and detect tiredness. Convolution neural networks utilizes extracted face features like yawning, eye flashing and head movements to detect exhaustion and sleepiness. Incorporating modified InceptionV3, VGG16, ResNet50, DenseNet201 and MobileNetV2 architecture over Driver Drowsiness Dataset to propose an ensemble deep learning model. Feature extraction was done using these models. The global max pooling layer is used to improve spatial robustness and dropout approach was included in these models to avoid overfitting on training data. Finally, Sigmoid classifier is used to classify positive (drowsy) or a negative (nondrowsy) result. These models outputs are given to a proposed ensemble algorithm. This model outperforms the alternative strategy with respect to performance metrics. The suggested ensemble framework performs better in identifying driver drowsiness than existing state-of-the-art techniques on basis of accuracy.