Driver alertness detection using CNN-BiLSTM and implementation on ARM-based SBC

Driver alertness detection is one of the significant automotive-related features to Advance Driver Assistance Systems (ADAS). Electroencephalogram based alertness detection is a direct method of determining consciousness level.  In this paper, an algorithm using a one-dimensional convolution neural network and bidirectional LSTM to learn the alertness level from EEG signals is proposed. The algorithm is implemented on an ARM-based single-board computer (SBC) for performance analysis. Real-time detection of drowsiness is necessary to alert the driver whenever he is about to sleep. Most of the existing methods focus on off-line analysis for interpreting the driver's state. The proposed method uses deep learning techniques to characterize and train the system, and the trained Model is ported to ARM SBC for real-time performance. Physionet sleep edf data with single-channel FPz-Cz is used for training the Model. The trained CNN-LSTM based Model gave an accuracy of 93.3% and the test model gave an accuracy of 89.4% when tested with real-time signals using the Neurosky mind wave electrode. To reduce road accidents occurring due to the driver's drowsiness, it is necessary to monitor driver alertness and alarm when necessary continuously. Driver alertness detection is one of the significant automotive-related features to Advance Driver Assistance Systems (ADAS). Electroencephalogram based alertness detection is a direct method of determining consciousness level.  In this paper, an algorithm using a one-dimensional convolution neural network and bidirectional LSTM to learn the alertness level from EEG signals is proposed. The algorithm is implemented on an ARM-based single-board computer (SBC) for performance analysis. Real-time detection of drowsiness is necessary to alert the driver whenever he is about to sleep. Most of the existing methods focus on off-line analysis for interpreting the driver's state. The proposed method uses deep learning techniques to characterize and train the system, and the trained Model is ported to ARM SBC for real-time performance. Physionet sleep edf data with single-channel FPz-Cz is used for training the Model. The trained CNN-LSTM based Model gave an accuracy of 93.3% and the test model gave an accuracy of 89.4% when tested with real-time signals using the Neurosky mind wave electrode. To reduce road accidents occurring due to the driver's drowsiness, it is necessary to monitor driver alertness and alarm when necessary continuously.