Detecting driver distraction using stimuli-response EEG analysis
URL: http://arxiv.org/abs/1904.09100v1
archive: archived pipeline: cataloged
Abstract
Detecting driver distraction is a significant concern for future intelligent transportation systems. We present a new approach for identifying a distracted driver's behavior by evaluating a stimulus and response interaction with the brain signals in two ways. First, measuring driver's response through EEG by creating various types of distraction stimuli such as reading, texting, calling and using phone camera (risk odds ratio of these activities determined by NHTSA study). Second, using a survey, comparing driver's order/perception of severity of distraction with the derived distraction index from EEG bands. A 14 electrodes headset was used to record the brain signals while driving in the pilot study with two subjects and a single dry electrode headset with 13 subjects in the main study. We used a naturalistic on-the-road driving study as opposed to a virtual-reality driving simulator to perform the distracted driving maneuvers, consisting of over 100 short duration trials (three to five seconds) for a subject. Machine learning methods achieved a mean accuracy (averaged over the subjects and tasks) of 91.54 +/- 5.23% to detect a distracted driving event and 76.99 +/- 8.63% to distinguish between the five distraction cases in our study (read, text, call, and snapshot) using a single electrode.
Summary
Naturalistic on-road EEG study examining whether common phone-related distractions (reading, texting, calling, taking a snapshot) produce detectable brain-signal patterns during real driving. A pilot used a 14-electrode Emotiv EPOC headset with 2 subjects; the main study used a single dry-electrode Neurosky Mindband with 13 subjects (15 total, ages 24-28 plus one 55-year-old, 9 male / 5 female). Trials were brief (60s recordings) on a closed campus road, with audio-prompted distractor tasks and synchronized video. EEG data were processed (filtering, ICA), analyzed in time and frequency domains, and classified using machine learning. Authors report localizing distraction-relevant activity to a single frontal electrode (FC5) and achieving 91.54 +/- 5.23% accuracy for distracted-vs-baseline detection and 76.99 +/- 8.63% for discriminating among the four distractor types. Authors propose a mobile real-time alerting application built on the single-sensor approach.
Key finding
On-road EEG recorded from a single dry frontal electrode (FC5) discriminated distracted from undistracted driving at 91.54% mean accuracy and distinguished four phone-related distractor categories (read, text, call, snapshot) at 76.99% mean accuracy across 15 subjects, supporting feasibility of low-channel EEG as a real-time distraction-detection signal in naturalistic driving.
Methodology
Naturalistic on-road experiment on a closed campus road (~1 mile, straight and curved). Pilot: 2 subjects with 14-electrode Emotiv EPOC (128 Hz). Main study: 13 subjects with single-electrode Neurosky Mindband (512 Hz). Within-subject design with five conditions per session: baseline (undistracted), reading printed text, texting, calling, and taking a phone snapshot, plus a 'pretend-to-text' control. Audio prompts cued tasks; sessions per subject ranged from 2 to 4. Synchronized video used for verification. Signal pipeline: noise filtering, independent component analysis, time- and frequency-domain feature extraction, then classification via standard machine-learning models with subject-level accuracy reporting.
Sample size: 15 subjects total (2 pilot with 14-electrode EPOC; 13 main with single-electrode Neurosky); 9 male / 5 female aged 24-28 plus one male aged 55; 2-4 sessions per subject.
Quality score: 5 / 5