Johns Hopkins AI predicts car crashes and tests safety fixes

Johns Hopkins University researchers have developed SafeTraffic Copilot, an AI tool that uses large language models to predict car accidents and assess how infrastructure changes will impact crash rates. The system can forecast how altering traffic light timing or other road conditions will affect safety outcomes, offering policymakers data-driven insights to reduce the 381 traffic fatalities that have occurred on Maryland highways this year alone.

How it works: SafeTraffic Copilot processes massive amounts of accident data using generative AI to understand complex crash scenarios and predict future risks.

• The model was trained on descriptions of more than 66,000 accidents, including road conditions, blood alcohol levels, satellite images, and on-site photography.
• Unlike traditional machine learning approaches, the AI can generate “what-if” scenarios, such as predicting accident changes if a traffic light timing shifts from 20 seconds to 30 seconds.
• The system provides confidence scores alongside predictions, addressing the “black box” problem that has previously deterred AI use in high-risk traffic safety applications.

Key findings: The research reveals alcohol and aggressive driving as the most dangerous factors, contributing to three times more crashes than other causes.

• Maryland fatalities have steadily risen over the past decade, from 466 deaths in 2013 to 621 in 2023.
• The AI can evaluate both individual and combined risk factors, offering more detailed understanding of how multiple elements influence safety outcomes.

What makes this different: SafeTraffic Copilot advances beyond current machine learning systems used by Maryland and other states for road safety assessment.

• “With machine learning, if a sample is not similar to your training samples, you cannot generate a prediction,” explained senior author Hao “Frank” Yang, an assistant professor of civil and systems engineering at Johns Hopkins. “Generative AI can give clear ‘what-if’ capabilities.”
• The system improves predictions with additional information and can be customized for different states or cities.

Global applications: The use of large language models allows the AI to adapt to traffic conditions in other countries and cultures as easily as providing a descriptive paragraph.

• “In South Asian countries like Taiwan or the Philippines, most crashes are motorcycle crashes, and the way they drive is different as well,” Yang said. “With previous models, incorporating drivers’ behavior or cultures is impossible.”
• Yang hopes to expand the research internationally while also benefiting local Baltimore and Maryland communities.

What they’re saying: “These are complex events affected by numerous variables, like weather, traffic patterns, roadway design and driver behavior,” Yang said. “With SafeTraffic Copilot, our goal is to simplify this complexity and provide infrastructure designers and policymakers with data-based insights to mitigate crashes.”