Trajectory Prediction for Multiple Agents in Dynamic Environments: Factoring in Traffic States and Driving Styles

Predicting the trajectories of multiple agents in dynamic driving scenarios such as intersections and roundabouts remains challenging due to dense agent interactions, varying speeds, and complex environmental constraints. While previous studies have proposed interaction-aware models, they often neglect the combined effects of traffic states and individual driving behaviors. This study presents a personalized motion prediction framework based on Vehicle-to-Vehicle (V2V) communication to support decision-making in complex environments. An unsupervised method is proposed to recognize driving styles by analyzing features derived from V2V communication, which are then encoded and incorporated into the prediction models, with feature importance analyzed via a random forest algorithm. Traffic states are modeled using relative spatial relationships and captured through spatiotemporal encoding. To explore different prediction paradigms, three architectures, GAN-based, CVAE-based, and Transformer-based, are designed and comparatively evaluated, each incorporating driving style and traffic state information as categorical encodings. Among these, the Transformer-based model demonstrates superior performance across multiple prediction horizons and metrics due to its effective integration of behavioral and contextual features in both encoder and decoder. The proposed framework is evaluated on two public interactive driving datasets, rounD and INTERACTION, where it achieves state-of-the-art accuracy. Results show that incorporating personalized driving styles and traffic context significantly enhances trajectory prediction, making it more suitable for real-time decision support in connected and autonomous vehicles.