Class-imbalanced crash prediction based on real-time traffic and weather data: A driving simulator study

Area of Science:

• Traffic Safety Engineering
• Machine Learning Applications
• Data-Driven Modeling

Background:

• Conventional statistical crash prediction models face challenges with data quality and volume.
• Machine learning (ML) algorithms offer potential but require careful parameter tuning for optimal performance.
• Real-time crash prediction is crucial for proactive traffic management and safety improvements.

Purpose of the Study:

• To develop and optimize real-time crash occurrence prediction models for traffic management systems.
• To compare the performance of Support Vector Machine (SVM) and Multilayer Perceptron (MLP) models in crash prediction.
• To identify key precursors of crashes within the driver-vehicle-environment framework.

Main Methods:

• Designed and optimized data-driven crash prediction models using SVM and MLP techniques.
• Assessed driver input responses, vehicle kinematics, and weather conditions from a driving simulator.
• Employed the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance in crash data.

Main Results:

• MLP demonstrated superior prediction performance, achieving over 94% recall in clear, overcast, and snow conditions.
• MLP and SVM achieved high F1-scores in various weather conditions (overcast, rain, snow).
• MLP and SVM obtained over 90% G-mean levels in fog, rain, and snow conditions, indicating robust performance.

Conclusions:

• The developed ML models offer significant insights into crash event forecasting.
• Findings support the design of enhanced crash avoidance/warning systems.
• Model effectiveness is validated based on driver input, vehicle kinematics, and diverse weather conditions.