Predicting collision cases at unsignalized intersections using EEG metrics and driving simulator platform.

Area of Science:

• Neuroscience
• Traffic Safety Engineering
• Machine Learning

Background:

• Unsignalized intersection collisions pose significant global traffic safety risks.
• Predicting potential collisions and identifying road hazards remain challenging.
• Existing methods often overlook driver's cognitive and physiological states.

Purpose of the Study:

• To investigate the feasibility of using electroencephalography (EEG) metrics for predicting road hazards.
• To develop and evaluate an improved neural network model for intersection collision prediction using EEG and driving behavior data.
• To enhance traffic safety by identifying at-risk drivers.

Main Methods:

• Collected EEG metrics and driving behavior data from drivers.
• Compared three machine learning models (MLP, LR, RF) using EEG, driving behavior, and combined datasets.
• Improved the MLP model with an attention mechanism and feature selection using Random Forest for multi-time point prediction.

Main Results:

• EEG metrics showed significant differences between collision and non-collision scenarios.
• Drivers with higher relative power in alpha and beta bands and lower power in delta and theta bands were more prone to conflicts.
• The improved MLP model with attention and feature selection achieved 88% accuracy in multi-time point prediction, outperforming baseline models.

Conclusions:

• EEG metrics are effective indicators for predicting collision probability.
• The enhanced neural network model demonstrates high accuracy in identifying unsafe drivers.
• This approach holds significant potential for reducing intersection accidents and improving overall traffic safety.