Files
Abstract
Effective prediction of traffic collision outcomes is essential for enhancing road safety and optimizingtraffic management. This thesis introduces a novel architecture aimed at improving collision
outcome predictions by utilizing a comprehensive dataset curated from multiple sources, including
weather, crash reports, traffic data, pavement geometry, and facility characteristics. The core of
this approach is the creation of a Tabular Transformer model that organizes disparate data into
distinct feature groups (traffic, event, vehicle, driver, environmental, geometric, pavement, contextual),
which are represented as tokens. These group-based tokens serve as essential, semantic rich
components for inferring collision patterns and deciphering causality. The study also conducts a
comparative analysis with widely used tree ensemble models to evaluate the efficacy of the proposed
method. The results illuminate the factors influencing various crash types, offering fresh
perspectives in the realm of tabular data utilization and result interpretation.