Using Data Augmentation to Improve EEG-Based Prediction of Emergency Braking Intention

Abstract

Accurate recognition of the driver's emergency braking intention can make the vehicle’ emergency braking system work automatically in advance, thus avoiding some serious traffic accidents. In this paper, we investigated the driver’s emergency braking intention detection based on EEG with data augmentation (DA). We used two electroencephalography (EEG) datasets of the driver’s emergency braking intention detection and performed DA on each of them. Five representative algorithms were used to detect the intention of emergency braking. The results showed that DA significantly improved the prediction accuracy of three convolutional neural networks (ShallowNet, DeepNet and EEGNet) based on deep learning, and slightly improved the prediction performance of the minimum distance to mean based on Riemannian geometry (RMDM) algorithm, but had little impact on the classical common spatial pattern filter combined with linear discriminant analysis (CSP+LDA) algorithm. With DA, EEGNet predicted the impending emergency braking behavior 200ms in advance with an accuracy of over 80% on both datasets. This study provides a framework for detecting drivers’ emergency braking intention based on EEG, which can improve the prediction accuracy by augmenting the EEG data with a small number of samples.