Exploring the Determinants of Pedestrian Crash Severity Using an AutoML Approach

ABSTRACT

            The increasing emphasis on pedestrian safety has prompted transportation engineers and urban planners to probe deeper into factors associated with pedestrian crashes. Recognizing this importance, the present study utilizes Automated Machine Learning (AutoML) to analyze 12 years of crash data from Utah spanning 2010 to 2022, where a total of 9,529 crashes occurred with pedestrian involvement, aiming to comprehend the multifaceted determinants of pedestrian-involved crash outcomes. These incidents are classified into severity levels: Fatal, Serious Injury, Minor Injury, Possible Injury, and No Injury / Property Damage Only (PDO). In analyzing potential predictors, the variables are categorized as demographics (sex and age group), driver behavior (aggressive driving, alcohol-drug test results, DUI, distracted driving, drowsy driving, and wrong way driving), particular crash scenarios (older or teenage driver involvement, crashes during holidays, involvement of right-turn, intersection, left-turn, overturn rollover, and presence of domestic animal), vehicle type and condition (involvement of commercial vehicles, heavy trucks, transit vehicles, and work zones), and environmental and infrastructure factors (road type, functional class, roadway surface condition, lighting condition, weather condition, and vertical alignment). Given the dataset’s rich context, with probable complex interactions and vast dimensionality, conventional modeling techniques liked ordered regression might fall short. This study utilizes AutoML to automate the machine learning process, from data preprocessing to model selection and tuning. AutoML explores a variety of algorithms, autonomously selects the optimal model based on accuracy, and fine-tunes its parameters for the given dataset. The approach ensures the model is robust, minimizes overfitting, and identifies key factors influencing pedestrian crash severity. By leveraging AutoML, the research streamlines traditional modeling efforts, achieving efficient and accurate results. Initial results indicate that the AutoML model achieved an accuracy of 86 percent in predicting crash severity. The findings underscore the profound influence of driver behavior, as well as environmental and infrastructure factors, on pedestrian-involved crashes. With a focus on the salient determinants of pedestrian crash severity, this study aims to contribute substantively to pedestrian safety research. The insights offer a strong foundation for shaping data-informed policies and infrastructural modifications, aspiring to foster safer pedestrian environments in Utah and potentially other similar regions.

Keywords: Pedestrian Safety, Crash Severity, Data-driven Policy, AutoML