Research on Vehicle Trajectory Prediction Based on an Improved GAT Model

Due to the complexity and high interactivity of the transportation environment, there is inherent uncertainty in the behavior of road participants, making real-time and accurate vehicle trajectory prediction a major challenge in the field of autonomous driving. This article proposes a novel multimodal vehicle trajectory prediction framework based on GAT form, which comprehensively considers motion uncertainty and the interactions between vehicles, lane markings, and road boundaries. This model introduces a novel improved graph attention mechanism and a time encoder to model the dynamic interaction information between multiple vehicles, and achieves spatiotemporal information fusion through gate units. In addition, we designed a Conditional Variational Encoder (CAVE) that considers the impact of road topology information (such as lane geometry and intersection layout) and traffic rules (such as speed limits and turning restrictions) on vehicle motion, and utilizes the fusion of spatiotemporal information to achieve multimodal prediction. A large number of experiments on the HighD dataset have shown that compared to state-of-the-art methods, the proposed model significantly reduces both short-term and long-term prediction errors, and can generate physically reasonable and diverse trajectory hypotheses. This method can provide more reliable decision support for auto drive system and traffic management applications.