Analysis of lane-changing primitives in highway merging areas based on nonparametric Bayesian models

Accurate parsing of lane-changing behavior in highway merging areas is crucial for understanding driving intent and ensuring traffic safety. Driving primitive segmentation has been recognized as an effective approach for interpreting complex lane-changing behaviors. However, the time-varying and interaction issues for driving primitive analysis are often ignored. This paper proposes a driving primitive segmentation method based on Hierarchical Dirichlet Process–Hidden Semi-Markov Model (HDP-HSMM), a nonparametric Bayesian framework adaptively infers primitive categories and durations from trajectory data. Using 1,330 naturalistic lane-changing trajectories in the merging area, the proposed approach is systematically compared with th Gaussian Mixture Model–Hidden Markov Model (GMM-HMM). Empirical findings demonstrate that HDP-HSMM yields more temporally consistent and semantically meaningful primitives. Specifically, the distribution of primitive durations under HDP-HSMM is concentrated within the 1.0–2.0 second range, to eliminate the results with excessively short fragments and overly long segments in GMM-HMM model. Moreover, the segmentation points identified by HDP-HSMM exhibit a bimodal distribution, peaking at relative positions of 0.3–0.4 and 0.7–0.8, which aligns well with the typical phases of lane-changing maneuvers. In contrast, GMM-HMM produces segmentation points concentrated mainly in the early phase (0.08–0.4) of the maneuver. These results confirm that HDP-HSMM is more effective on capturing the intrinsic structure of driving behavior, and providing a more reliable basis for behavior analysis and decision-making in intelligent transportation systems.