ECG-Based Driver Distraction Identification Using Wavelet Packet Transform and Discriminative Kernel-Based Features

Abstract

Driver Distraction is one of the main reasons behind the increasing number of fatalities on the road. In order to minimize the potential road disasters, it is essential to monitor and track the pre-requisites of driver distraction. While driving, the driver might get distracted in variety of ways such as talking on the cell phone, texting, or having a conversation with a passenger. In the recent years, extensive investigations are directed towards the problem of characterizing the impact of secondary tasks while driving, predominantly using camera-based systems. However, camera-based systems incur major challenges such as privacy or latency in detection. Using physiological signals to identify distraction such as Electroencephalography (EEG) has been shown to accomplish more reliable detection. However, EEG- based detection systems necessitate intrusive implementation and complex signal processing. On the other hand, Electrocardiogram (ECG) is a reliable signal which can characterize the physiological changes consistently, with minimal intrusiveness, and at low cost. In this paper, we focus on ECG signal processing aspect with the aim of predicting driver distraction. Eight subjects aged 24 ± 45, actively participated in the naturalistic driving experiment where distraction was induced by: 1) phone conversation and 2) engaging an active conversation between the driver and the passenger. We present an effective frequency subBand analysis using Wavelet Packet Transform (WPT) to localize the impact of distracting elements. Due to high dimensionality of the WPT generated space, we then applied Linear Discriminant Analysis (LDA) for feature space dimensionality reduction; preserving discriminative capability of the predictive model. In order to further enhance the prediction ability of the system, we used kernel transformation in order to take into account non-linear interactions of the input feature space. Based on our results, WPT transform in combination with Linear Discriminant dimensionality reduction demonstrated high potentials to detect normal vs. distracted driving scenarios. Using kernel transformation further increased feature space discrimination compared to the baseline features and let to an increase from 44.10% to 88.45% average prediction accuracy over all subjects.