Can we realize seamless traffic safety at smart intersections by predicting and preventing impending crashes?

Area of Science:

• Traffic Safety Engineering
• Artificial Intelligence in Transportation
• Urban Planning

Background:

• Road intersections are major contributors to urban traffic accidents and casualties.
• Existing crash prediction models lack effectiveness in capturing intersection-specific complexities like varied crash types and cyclical traffic patterns.
• Previous research often relies on data sampling for imbalance, overlooking advanced anomaly detection methods.

Purpose of the Study:

• To develop a reliable model for predicting intersection-level crash likelihood.
• To address the limitations of existing models in accounting for Signal Phasing and Timing (SPaT) variations and traffic flow dynamics.
• To explore advanced anomaly detection techniques for imbalanced crash data.

Main Methods:

• An anomaly detection framework integrating Generative Adversarial Networks (GANs) and Transformers was developed.
• The model utilizes high-resolution event data from Automated Traffic Signal Performance Measures (ATSPM), including SPaT and traffic flow.
• Data was sourced from 11 intersections in Seminole County, Florida, focusing on cycle-level crash prediction.

Main Results:

• The proposed framework achieved 76% sensitivity in predicting crash events.
• The model effectively handles highly imbalanced crash data by integrating real-world SPaT and traffic insights.
• Demonstrated potential for real-time application in smart intersection environments.

Conclusions:

• The developed GANs-Transformers framework offers a promising solution for proactive intersection crash prevention.
• The study provides a viable roadmap for city-wide implementation in smart cities, enhancing traffic safety.
• Potential real-time applications include adaptive signal timing, driver warnings, and optimized emergency responses, contributing to safer urban mobility.