Interactive Trajectory Primitives Representation and Extraction Based on Naturalistic Driving Data

In shared road space，there are path conflicts between different road users moving in various directions. Road users must negotiate right-of-way through driving interactions to avoid collision risks，thus resolving potential conflicts. The description and modeling of interactive behaviors is crucial for accurately understanding and predicting the dynamic environment. Therefore，a semantic-level representation and extraction method for multi-vehicle interactive behaviors is proposed in this paper，taking interactive trajectory primitives as analysis units. Firstly，a nonparametric Bayesian method is utilized to segment interactive behaviors，obtaining interaction segments with significant behavior patterns. Then，the sticky hierarchical Dirichlet-Hidden Markov Model is employed to extract interaction primitives from these interaction segments. Finally，unsupervised clustering is applied to the normalized interaction primitives to obtain semantic-level behavioral features of interaction scenarios. An empirical study based on 20 797 pairs of multi-vehicle interaction data from the NGSIM highway dataset shows that the method proposed in this paper can extract and analyze complex interactive scenarios involving multiple participants，breaking through the limitation of existing research that only constructs interaction primitives for two vehicle interaction scenarios，and supporting the analysis of interaction among multiple traffic participants. The experimental results show that the proposed method can segment continuous driving behaviors into discrete interaction primitives. The clustering results correspond to actual interaction scenarios and can be used to characterize the interaction behaviors among vehicles in different interactive trajectory primitives. Furthermore，the method can enhance performance of downstream driving tasks in complex scenarios. In multi-step vehicle trajectory prediction，by integrating with baseline prediction methods，the proposed method can reduce the average prediction error and final position error by 19.3% and 14.6%，respectively，compared to baseline methods. © 2024 SAE-China. All rights reserved.