1. Home
  2. RIP
  3. View Record
https://www.nationalacademies.org/trb/events

Development of AI/ML Algorithms to Extract Information from Crash Report Diagrams for Statewide Left Turn-Related Crashes at Intersections

Left turn-related crashes at intersections remain one of the most prevalent and dangerous types of motor vehicle collisions in North Carolina and across the United States. These crashes are inherently complex, shaped by conflicting vehicle paths and diverse intersection geometries and can overwhelm a driver’s attention and decision-making processes. In 2023, nearly one-third of all fatal and serious crashes in North Carolina were attributed to left turn maneuvers, underscoring a critical and persistent safety concern. Yet, despite their significance, left turn-related crashes are routinely misclassified in police-reported data and poorly captured by existing structured datasets, obscuring the true scale and characteristics of the problem. The proposed research addresses this longstanding data gap by developing and validating a novel AI/ML framework to extract high-value safety information from crash diagrams submitted as part of electronic crash reports. By leveraging convolutional neural networks and multi-task learning, this effort will automate the classification and characterization of left turn-related crashes at intersections—identifying key features such as intersection type, crash location, turning movements, lane configurations, and physical site elements (e.g., medians, islands). These details, often embedded in crash diagrams but absent from structured data, are essential for diagnosing intersection safety problems and formulating effective countermeasures. The research will proceed through a comprehensive, multi-phase methodology. A foundational element of this effort is the manual review and annotation of over 1,000 crash diagrams to develop a robust training and validation dataset. Using this dataset, the project team will train and test AI/ML models to accurately detect whether a crash occurred within an intersection, whether it was left turn-related, and extract additional contextual features from the diagram. This research fills a critical void in the current traffic safety analysis paradigm by automating a previously manual and time-intensive process. It also represents one of the first large-scale applications of AI to the interpretation of unstructured visual crash data—transforming static crash diagrams into actionable intelligence. This project supports both the North Carolina Department of Transportation’s (NCDOT) Strategic Highway Safety Plan (2024) and Highway Safety Improvement Program (2024), which prioritize intersection safety and seek to halve traffic-related fatalities and serious injuries by 2035. Upon completion, the research will yield a curated dataset of left turn-related intersection crashes, and a framework for scalable deployment across other crash types or jurisdictions. This initiative holds significant promise not only in improving the fidelity of crash data but also in enhancing the efficiency, accuracy, and responsiveness of safety analyses at the state level. The methodology developed here will serve as a foundational model for future research and applications aimed at other high-priority crash types—ushering in a new era of data-driven, AI-enabled traffic safety management.