A similarity-based data-driven car-following model considering driver heterogeneity

Abstract

Human drivers usually have distinct driving patterns and preferences. Driver heterogeneity is crucial for modeling driving behaviors. This paper incorporates driver heterogeneity with data-driven approaches to predict car-following behaviors. A bi-level similarity-based car-following model is proposed to predict the vehicle's moving distance. In the upper level, drivers with similar driving patterns as the ego vehicle are identified using k-nearest neighboring (kNN) search. In the lower level, leveraging kNN model, candidate records are selected from the identified vehicles’ trajectories and applied to predict the ego vehicle's moving distance, combining the driving pattern similarity measured in the upper level. By taking into account the driver heterogeneity, the proposed model is capable of identifying the most relevant driving situations, which leads to an improvement of prediction accuracy. Furthermore, the established bi-level structure largely shrinks the searching space of candidate records, which reduces the searching complexity and enhances computational efficiency. We quantitatively evaluate and compare the performance of the proposed model in terms of both prediction accuracy and computational efficiency using real-world vehicle trajectory data.