Driver vigilance estimation with Bayesian LSTM Auto-encoder and XGBoost using EEG/EOG data

Abstract

Driver drowsiness remains one of the main causes of road accidents even if solutions based on machine learning (ML) applied to frontal Electroencephalogram (EEG) and Electroocoulogram (EOG) data makes it possible to detect and prevent the driver's fatigue level. In this paper, we bring forward two different variational auto encoders (VAE) models for the estimation of driver alertness. Our proposed approach aims to reduce the large amount of EEG features and generate meaningful EEG features while preserving a reasonable level of accuracy in a classifier. To demonstrate the validity of our proposed method, our experiments were conducted using a multimodal publicly available dataset using only EEG and EOG forehead features. After the projection, those features are applied to an XGboost classifier to predict the attention state. We demonstrate that the long short-term memory (LSTM) encoder provides a better encoding process besides that the bayesian version of the LSTM auto encoder provides almost the same significant extraction. VAE LSTM can be used for both feature extraction and anomaly detection in one framework.