Heterogeneity analysis of car-following behavior using two-stage unsupervised learning method

In highway congestion scenarios, traffic accidents are frequently influenced by drivers’ longitudinal maneuvers and interactions between vehicles. To enhance the longitudinal control safety of autonomous vehicles, this study aims to extract car-following behavior patterns in typical Chinese highway congestion scenarios and analyze the distribution differences and risk characteristics of these patterns across different vehicle types. A two-stage unsupervised learning method is proposed: first, car-following trajectories were segmented by CHMM algorithm. Second, GMM-LDA algorithm is applied to cluster these segments and identify distinct behavior patterns. The results demonstrate that the two-stage method effectively describes the operational and interactive states of vehicles and reveals significant differences in car-following behavioral patterns across vehicle types and driving risk levels. This study provides data-driven prior knowledge for the longitudinal control of autonomous vehicles and offers new insights to understand the heterogeneity of car-following behaviors.