Geometry-aware car-following model construction: Theoretical modeling and empirical analysis on horizontal curves

Abstract

Road geometry significantly influences the physical forces acting on vehicles and the perceptual ability of drivers. Unfortunately, most available car-following models ignore the influence of complex road geographical features, such as curvatures and slopes and thereby lack scalability. To fill these gaps, this study presents a framework for the construction of a geometry-aware car-following model. Under the over-alignment assumption, car-following motion on horizontal curves was simplified into seven internal or adjacent car-following scenarios. Two novel alternative vehicle control modes (centralized and decentralized) for car-following motions on a horizontal route were proposed. The structured features of each scenario, considering both lateral and longitudinal information, were defined mathematically. Open-source data with trajectory records and road surface conditions on highways in Japan were collected and used as empirical data sources. First, we analyzed the theoretical proportion of traffic scenarios that conformed to the traditional car-following model for any horizontal route. Several properties of the car-following scenario proportion were proposed and proved. Both empirical statistics and theoretical estimations showed the existence of real-world sizable car-following scenarios that could not be handled by traditional models. Owing to their powerful ability to handle complex input features, machining learning and deep learning models were applied in car-following behavior modeling to make multistep predictions. With high computational efficiency, the results were compared with those of models with traditional inputs to demonstrate the effectiveness of the proposed approach.