An integrated methodology for real-time driving risk status prediction using naturalistic driving data

Area of Science:

• Traffic Safety
• Artificial Intelligence
• Transportation Engineering

Background:

• Real-time driving risk status prediction is vital for proactive traffic interventions and enhancing road safety.
• Optimal observation and prediction time window lengths for these models are underexplored.
• Existing methods often lack the precision needed for timely and accurate risk assessment.

Purpose of the Study:

• To propose and validate a methodology for accurate real-time driving risk status evaluation and prediction.
• To determine optimal time window lengths for observation and prediction in driving risk assessment.
• To identify key influencing factors affecting driving risk status.

Main Methods:

• Developed a methodology integrating driving risk status identification, rolling time window feature extraction, and prediction.
• Utilized a dataset of 1,440 car-following events from the Shanghai Naturalistic Driving Study.
• Employed a multi-layer perceptron model for risk status prediction and analyzed feature importance.

Main Results:

• Identified four driving risk statuses (safe, low-risk, median-risk, high-risk) as optimal for risk labeling.
• Achieved over 85% accuracy in predicting medium-risk or high-risk status within 0.7 seconds using a 0.5-second observation window.
• Confirmed that variables related to risk status score, speed difference, headway distance, speed, and acceleration are significant predictors.

Conclusions:

• The proposed methodology effectively evaluates and predicts real-time driving risk status.
• Optimal time window lengths (0.5s observation, 0.7s prediction) significantly improve prediction accuracy.
• Findings support the application of these methods in connected and autonomous vehicles (CAVs) to reduce driver workload and enhance safety.