SafeTraffic Copilot is 70% accurate at predicting car crashes

Researchers have developed an artificial intelligence-based tool that can predict car crashes. The model could potentially reduce one of the leading causes of death in the United States.

According to a new study published in Nature Communications, the tool, called SafeTraffic Copilot, can provide a 70% accurate prediction in a real-world scenario.

The Johns Hopkins University researchers used large language models (LLMs) to train SafeTraffic Copilot using descriptions of road conditions, numerical values like blood alcohol levels, satellite images, and on-site photography.

"By reframing crash prediction as a reasoning task and using LLMs to integrate written and visual data, the stakeholders can move from coarse, aggregate statistics to a fine-tuned understanding of what causes specific crashes," senior author Hao (Frank) Yang, a professor of civil and systems engineering, said in a press release.

The authors say the model gives policymakers and transportation designers a reliable tool for identifying combinations of factors that raise the risk of car crashes.

However, they say the model shouldn't replace humans, who will remain the final decision-makers. Instead, it could serve as a pilot for processing information, identifying patterns, and quantifying risks.

More than 120 people die in crashes every day in the US, making it one of the leading causes of death, the CDC data shows.

Deaths from car crashes in 2022 resulted in over $470 billion in total costs, including medical expenses and cost estimates for lives lost.

AI has been increasingly used in road safety, as it allows for real-time traffic monitoring and driver behavior detection.

Just last week, Cambridge Mobile Telematics launched StreetVision, a platform that predicts road risks based on phone distraction, speeding, hard braking, and aggressive cornering, among other indicators.

However, critics say overreliance on predictive AI is common, and the lack of double-checking of AI predictions may lead to unnoticed errors.