A stacked ensemble model for traffic conflict prediction using emerging sensor data

Abstract

Over recent decades, a plethora of Safety Surrogate Measures (SSMs) have emerged as valuable metrics for predicting traffic conflicts. However, existing research predominantly focuses on conflict prediction at junctions or relies on a single SSM, such as time-to-collision, for detecting vehicle-related conflicts. This limitation highlights the challenge of selecting appropriate SSMs for vehicle- or segment-based conflict prediction, considering the diverse range of factors influencing conflict outcomes. To address this gap, this study leverages data collected from various infrastructure and vehicle-based sensors, including drones, lidars, radars and cameras, across different scenarios in China and the UK: urban junctions, motorway segments and vehicle-based data from instrumented vehicles. Employing machine learning approaches to handle the extensive and disaggregated data, a novel stacked ensemble learning model is proposed. This model integrates a Random Forest (RF), three-layer Deep Neural Networks (DNN), Support Vector Machine Radial (SVM-R), and a Gradient Boosting Model (GBM) meta layer to enhance prediction accuracy. The Recursive Feature Elimination (RFE) algorithm is then employed to identify the most influential SSMs for conflict prediction in each scenario. Results demonstrate the superiority of the stacked ensemble learning model, achieving accuracies of 88 % for junctions, 87.5 % for motorway segments and 99 % for vehicle-based conflicts. Furthermore, the study highlights the necessity of employing different SSMs for conflict detection in various scenarios. These findings hold significant implications for roadway operators and vehicle manufacturers, aiding in the development of strategies to detect infrastructure and vehicle-related traffic conflicts.