Vision-AI-driven dilemma zone modelling for driver behaviour prediction

Abstract

Driver behaviour in the dilemma zone critically influences red-light violations and intersection safety. This study presents a scalable Vision-AI and deep learning framework modelling driver decisions based on vehicle type, speed, acceleration, lane occupation, and signal distance. Validated at a signalised intersection in Thessaloniki, Greece, UAV-collected data yielded over 11,000 vehicle trajectories annotated with stop/go decisions during the yellow phase. A multi-layer perceptron classifier trained on this data reached 92.33% test accuracy and 91.76% F1-score. An ablation study analysed feature impacts, showcasing the model’s robustness and potential to improve intersection safety and inform intelligent traffic management.