Learning Pedestrian Failure-to-yield Maneuver Patterns from Fatal Crash Data: Evidence from Explainable AutoML

Abstract Pedestrian failure-to-yield crashes pose a growing threat to urban safety, with intersection-related violations contributing substantially to rising pedestrian fatalities in the United States. To investigate the contextual and behavioral factors underlying these incidents, this study analyzes crash-level data from the 2016--2023 Fatality Analysis Reporting System. Using the AutoGluon Tabular AutoML framework, this study developed an ensemble classification model to predict pedestrian crossing maneuvers and employed SHapley Additive exPlanations for interpretability. Pedestrian maneuvers were categorized following Pedestrian and Bicycle Crash Analysis Tool (PBCAT) definitions, including Crossing from Left, Crossing from Right, Crossing with Unknown Direction, Parallel Path Same Direction, Parallel Path Opposite Direction, and Stationary. The best-performing model, LightGBM, achieved a test accuracy of 79\%, with crash year, pedestrian position, road functional class, and driver age emerging as key predictors. SHAP analysis revealed that Crossing from Left crashes are strongly influenced by pedestrian non-compliance and position at the time of crash, while Crossing from Right crashes are associated with older drivers, limited lateral scanning, and sight-line obstructions at slip lanes. Crossing with Unknown Direction maneuvers frequently involve midblock jaywalking, particularly among older pedestrians. These findings support targeted countermeasures, including improved crosswalk visibility, age-specific driver programs, and intersection redesign.