Stacked Ensemble Model for Traffic Conflict Prediction

This technical report presents a stacked ensemble model designed for predicting traffic conflicts across various road environments using multi-modal sensor data. The study addresses the critical issue of traffic safety, particularly in light of the significant number of traffic collisions reported globally. The primary objectives include developing an accurate prediction model and assessing the transferability of this model across different scenarios, such as urban junctions and high-speed roadways. The methodology involves data acquisition from various sensors, including drones and instrumented vehicles, and the application of sophisticated computer vision techniques to process this data. The report details the selection of thirteen safety surrogate measures (SSMs) and the implementation of a stacked ensemble learning approach that combines multiple machine learning models to enhance prediction accuracy. Findings indicate varying effectiveness of SSMs in different contexts, with the vehicle-based scenario achieving the highest accuracy. The results underscore the importance of context-specific applications in traffic safety management.