Real-driving-implementable drowsy driving detection method using heart rate variability based on long short-term memory and autoencoder

Abstract

Abstract Drowsy driving is a fatal problem that may cause traffic accidents. Although many driver drowsiness detection methods have been proposed, most of them have problems in input data availability, robustness to real driving environments, or detection precision. A drowsiness detection method based on heart rate variability (HRV), which is an R-R interval (RRI) fluctuation obtained from an electrocardiogram (ECG), has been proposed since ECG is easy to measure by using a wearable sensor. HRV is related to the autonomic nervous system (ANS) and is affected by drowsiness. However, its drowsiness detection performance was not always satisfactory. This study proposes a new drowsiness detection method using raw RRI data instead of HRV to improve the drowsiness detection performance. The proposed method uses raw RRI time series as inputs, and a drowsiness detection model is trained based on long short-term memory (LSTM) and autoencoder (AE), which are types of neural networks. RRI data during driving were collected from 25 participants using a driving simulator. The drowsiness detection model was trained following the proposed method. The experimental result showed that the proposed method achieved an AUC of 0.88, a sensitivity of 81%, and a specificity of 91%, which was higher than the HRV-based method. The result suggests that it is better to use raw RRIs as inputs than HRV features.