Bayesian network for red-light-running prediction at signalized intersections

Abstract

Red-light-running (RLR) is an important reason for the large number of intersection-related fatalities, injuries, and other losses. The accurate RLR prediction can effectively reduce crashes caused by RLR behavior. The RLR prediction is usually composed of two parts: the vehicle’s stop-or-go behavior and the arrival time when the vehicle reaches the stop line. Previous stop-or-go prediction models are usually based on embedded traffic sensors using machine learning algorithms. While based on the continuous trajectories collected by radar sensors, RLR prediction can be conducted more effectively. In this paper, a probabilistic stop-or-go prediction model based on the Bayesian network (BN) is proposed for RLR prediction. We extend the deterministic output into the probabilistic output, which provides decision-makers with greater autonomy. The causality of BN improves the interpretability of the prediction model. The BN model is calibrated and tested by the continuous trajectories data measured by radar sensors installed at a signalized intersection. We not only consider the movement measurements of individual vehicles (e.g., speed and acceleration), but also take into account the car-following behavior. As a comparison, different machine learning models and the model based on the inductive loop detection (ILD) are adopted. The results show that the proposed BN model has a high prediction accuracy and performs better in the feature interpretation. This paper provides a new way for probabilistic RLR prediction based on continuous trajectories, which will significantly improve traffic safety of signalized intersections.