ENHANCING VULNERABLE ROAD USER SAFETY USING MACHINE LEARNING AND CROWDSOURCED DATA: A STUDY OF PEDESTRIAN CRASHES AT SIGNALIZED AND NON-SIGNALIZED INTERSECTIONS

The number of vehicle-pedestrian crashes is increasing nationally and globally. Further, due to their lack of physical protection, pedestrians typically sustain severe injury when a crash occurs. There exists a need to identify the factors that affect pedestrian crashes frequency and severity, and to examine how these factors are different at signalized and non-signalized intersections. Therefore, this dissertation presents a comprehensive approach to investigating pedestrian safety (frequency and severity of crashes) at signalized and non-signalized intersections. Also, by identifying factors associated with higher risk of pedestrian crashes, this dissertation addresses the adequacy of the existing signalization warrants in practice.

The study dataset combines crash data, pedestrian and vehicle volumes, and land use data. To streamline the data collection of intersection features, a software was developed in this study, reducing the time required by threefold. The data contains emerging crowdsourced and mobile-based data to capture pedestrian volumes. Negative binomial and ordered logit models were for the model calibration.

Regarding crash severity, the study reveals that both driver and pedestrian impairment, multilane roads, and the presence of clear zones are significant factors of pedestrian crashes at both intersection types. However, at signalized intersections, proximity to a college campus and the presence of push-button devices are associated with less severe outcomes, while nighttime conditions significantly increase crash severity. At non-signalized intersection, the absence of lighting infrastructure during nighttime contributes to more severe crashes.

A key methodological contribution is the hybrid approach developed to correct misclassified pedestrian crashes by integrating structured crash data with unstructured narrative reports. This method combines manual and semi-automated processes with natural language processing to accurately classify crash severity, identifying and reclassifying 5.5% of crashes in the study. The study provides a comprehensive comparative analysis of pedestrian crashes at signalized and non-signalized intersections, offering valuable insights for urban planners, traffic engineers, and policymakers in developing safer intersection designs and implementing data-driven safety interventions across diverse urban environments.