A hybrid framework integrating cellular automata and XGBoost for real-time traffic accident prediction at signalized intersections

Traffic accidents at signalized intersections constitute a major public safety concern, representing a significant proportion of urban road fatalities worldwide. Machine learning methods have been applied to accident prediction; however, most existing approaches depend on large-scale real-world datasets, focus on freeway segments rather than intersections, and operate at spatial and temporal resolutions that constrain their applicability for real-time event-level prediction. This paper presents a novel framework integrating a Cellular Automata (CA) traffic simulation model with an XGBoost (eXtreme Gradient Boosting) binary classifier for individual-event accident prediction at signalized intersections. The CA model enables the controlled generation of feature-rich data including heterogeneous vehicular dynamics and red-light running. The framework employs a sliding temporal window to extract features from simulation data for a 30-second prediction horizon. The proposed model achieves an accuracy of 0.86, a recall of 0.966, and an AUC of 0.91 on the test set, outperforming baseline models such as K-Nearest Neighbors and Long Short-Term Memory networks. The high recall is particularly important in safety-critical applications, as it minimizes undetected accident events. The results show that integrating simulation-based traffic modeling with machine learning enables efficient accident prediction at intersections and offers a viable alternative to sensor-based approaches for intelligent traffic management systems.