Efficient mapping of crash risk at intersections with connected vehicle data and deep learning models

Area of Science:

• Transportation Engineering
• Data Science
• Artificial Intelligence

Background:

• Traditional road safety assessments rely on historical crash data, requiring extensive collection periods (3+ years).
• Emerging connected vehicles (CVs) technology offers real-time data for proactive identification of hazardous road sections.
• Sharing intersection risk information via CVs can promote safer driving and guide infrastructure improvements.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for predicting intersection risk levels using CVs data.
• To compare the performance of deep learning models (MLP, CNN) against a traditional decision tree model.
• To assess the feasibility of using CVs data for efficient and accurate road safety analysis.

Main Methods:

• Utilized one month of CVs data from the Michigan Safety Pilot program and historical traffic/crash data for 774 intersections in Ann Arbor, Michigan.
• Extracted 24 features from both CVs and traffic data for each intersection.
• Trained Multi-Layer Perceptron (MLP) and Convolutional Neural Network (CNN) models, optimizing hyperparameters with Bayesian optimization.

Main Results:

• Deep learning models achieved accuracies over 90%, surpassing the decision tree model's ~87% accuracy.
• The CNN model demonstrated the highest accuracy at 93.8%, indicating its capability to capture complex data patterns.
• Interpretability analysis confirmed the validity and reliability of the CNN model's predictions.

Conclusions:

• The combination of CVs data and deep learning networks (MLP, CNN) is a promising approach for high-efficiency, low-penetration rate crash risk determination at intersections.
• This method enables proactive identification of high-risk intersections, facilitating timely deployment of countermeasures.
• The study highlights the potential of CVs technology and AI to significantly enhance traffic safety and reduce crash rates.