A novel deep learning-based technique for driver drowsiness detection

Every year, many people lose their lives because of road accidents. It is evident from statistics that drowsiness is one of the main causes of a large number of car accidents. In our research, we wish to solve this major problem by measuring the drowsiness level of the human brain while driving. The study aims to develop a novel technique to detect different alertness levels (i.e., awake, moderately drowsy, and maximally drowsy) of a person while driving. A hybrid model using a stacked autoencoder and hyperbolic tangent Long Short-Term Memory (TLSTM) network with attention mechanism is designed for this purpose. The designed model uses different biopotential signals, such as electroencephalography (EEG), facial electromyography (EMG), and different biomarkers, such as pulse rate, respiration rate galvanic skin response, and head movement to detect a person's alertness level. Here, the stacked autoencoder model is used for automated feature extraction. TLSTM is used to predict a person's alertness level using stacked autoencoder network-extracted features. The proposed model can classify awake, moderately drowsy, and maximally drowsy states of a person with accuracies of 99%, 98.3%, and 98.6%, respectively. The novel contributions of the paper includes (i) incorporation of an attention mechanism into the TLSTM network of the proposed hybrid model to focus on the emphatic states to enhance classification accuracy, and (ii) utilization of EEG, facial EMG, pulse rate, respiration rate, galvanic skin reaction, and head movement pattern to assess a person's alertness level.