Trajectory Prediction Based on Spatiotemporal Features of Multi-Source Directed Graphs

Trajectory prediction is a crucial component of autonomous driving systems, and accurate modeling of driving scenarios serves as the foundation for achieving reliable motion prediction. In recent years, scene modeling and motion prediction have garnered increasing attention from experts and scholars in the field. However, current trajectory prediction methods are constrained by the insufficient representation of dynamic interaction information in scene models and the lack of comprehensive consideration of map information in prediction models, which hinders the ability to capture multimodal characteristics and ultimately affects the accuracy of the predictions. To address these limitations, this paper focuses on scene modeling and multimodal trajectory prediction for autonomous vehicles. By constructing a multi-source heterogeneous directed graph, high-precision map information and vehicle-to-vehicle interaction features are extracted. Meanwhile, a GAT-Transformer network is developed to capture temporal dependencies and spatial position characteristics, enabling the generation of multimodal predicted trajectories for vehicles along with their probability distributions.