Recognizing drowsiness in young men during real driving based on electroencephalography using an end-to-end deep learning approach. (August 2021)

• Record Type: Journal Article
• Authors: Chen, Jichi; Wang, Shjie; He, Enqiu; Wang, Hong; Wang, Lin
• Is Part Of: Biomedical signal processing and control. Volume 69(2021)
• Publication Date: 2021-08
• Subjects: Automatic detection -- Convolutional neural network (CNN) -- Electroencephalography (EEG) -- Deep learning (DL) -- Driver drowsiness; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102792
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: To the best of our knowledge, so far deep learning theory and convolutional neural network is first applied on raw multi-channel EEG signals for driver drowsiness detection during real driving. We employ data augmentation strategy to increase the amount of training data, which is crucial for the successful application of deep ConvNets. The 12-layer deep ConvNets model achieves an accuracy, precision, sensitivity, specificity, and mean f-measure of 97.02 % ± 0.0177, 96.74 % ± 0.0347, 97.76 % ± 0.0168, 96.22 % ± 0.0426, and 97.19 % ± 0.0157, respectively. Abstract: It is widely agreed that driving while drowsy is a severe threat to road safety. Therefore, in this work, we present a novel approach that does not require manual selection of feature sets and then delivers them to the classifier, using deep learning theory and convolutional neural network (ConvNets) to automatically detect driver drowsiness based on multi-channel EEG signals during real driving. The proposed 12-layer deep ConvNets model automatically learns and extracts the most prominent features from the raw EEG data through 5 convolutional layers, 3 max pooling layers and 1 mean pooling layer and optimizes the classification results through 3 fully connected layers at the same time, which is an end-to-end manner. To overcome the lack of a large amount of EEG data, a data augmentation strategy is proposed. The proposed deep ConvNets model is trained on 4 s segments of EEG data from different participants and tested using a 10-fold cross validation. It gave an accuracy, precision, sensitivity, specificity, and mean f-measure of 97.02 % ± 0.0177, 96.74 % ± 0.0347, 97.76 % ± 0.0168, 96.22 % ± 0.0426, and 97.19 % ± 0.0157, respectively on the testing data set and outperforms the state-of-the-art systems, which proved the good generalization performance of the deep model. Considering that the proposed model can learn features from the data without using specialized feature extraction and classification methods, ConvNets may be considered as an alternative for similar detections based on EEG signals such as operators fatigue in navigation, construction industry, etc.