Abstract

The driver's mental state is frequently detected employing EEG signals which are usually converted into grayscale images to train a Machine Learning algorithm that classifies his mental status. This work aims to achieve a simplified and accurate method to detect the emergency braking intention employing EEG signals and a Convolutional Neural Network (CNN). Three main problems: computer resources, network accuracy, and the training time are defined to accomplish this aim. While a CNN is an efficient image-based classifier, it increases computing resources and processing time. Therefore, we solved these problems by training a CNN through a 2D matrices tensor designed to work with a very large database without transforming the EEG signals into grayscale images and running on a free cloud platform. However, we are well aware that physical fatigue while driving increases the mental load. Consequently, we measured the braking reaction time that proves an increment over time, negatively affecting the participants' performance. The linear correlation between the target and non-target classes on the matrices tensor reveals that most emergency events can be very well-differentiated from not anomalous driving. The CCN accuracy is over 84% with just four electrodes-scalp, comparable to reported grayscale-based methods.