Causal effects of road geometry on car-following behaviour of autonomous vehicle pairs and human-driven vehicle pairs

This study adopted a meta-learner (X-learner) in machine learning to investigate the causal effects of curvature and slope on car-following behaviour. The effects were analyzed using real-world trajectories of human-driven vehicles (HDVs) and autonomous vehicles (AVs) in different road geometric conditions in Italy, Sweden, and Hungary. The study considered the effects of curvature and slope at different look-ahead distances to capture the drivers’ anticipatory behaviour before entering curves and slopes. These effects were predicted using two machine learning models—eXtreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine. The results show that AVs and HDVs showed different speed patterns across look-ahead distances on curves and slopes, with notable differences between free-flow and car-following states. The results also showed that XGBoost and LightGBM models with the road geometry variables showed higher prediction accuracy than the models without the road geometry variables.