Car-following behavior is critical for traffic management, road design, and the development of advanced driver assistance systems (ADAS) and autonomous vehicles (AVs). Traditional theory-based car-following models are widely used in traffic simulations but rely on simplified assumptions, limiting their ability to capture the complexity of real-world driving. In contrast, machine learning (ML) models can leverage large datasets to uncover complex driving behaviors. However, a major limitation of ML models is their lack of interpretability. Moreover, the rise of AVs has introduced mixed-traffic environments where AVs and human-driven vehicles share the road. Understanding different interaction scenarios—such as AVs following human drivers (AH), human drivers following AVs (HA), and human drivers following other humans (HH)—is essential for accurate modeling and safe AV deployment. To address these challenges, we propose a car-following modeling framework that integrates the CatBoost algorithm with SHapley Additive exPlanations (SHAP). CatBoost handles both numerical and categorical data, enabling the development of scenario-specific models (AH, HA, HH) and a unified car-following model incorporating scenario type as a feature. SHAP enhances interpretability by quantifying the contribution of each model feature, e.g., speed and inter-vehicle distance, across scenarios. We apply this framework to the Lyft Level-5 dataset to analyze feature importance and evaluate how scenario type moderates driving behavior. The insights derived from our analysis support the design of more adaptive AV control strategies and inform transportation policies for the safe integration of AVs into modern traffic systems.
扫码关注我们
求助内容:
应助结果提醒方式:
