Multiple Graph Neural Networks and Transformers for Vehicle Trajectory Prediction

In complex traffic environments, it is essential for autonomous vehicles to plan their routes effectively by considering the actions of nearby road users. To better address this challenge, we introduce a new model for vehicle trajectory prediction called MGNformer, which combines multiple Graph Neural Networks and Transformers. The MGNformer encoder is structured to extract spatiotemporal features through two essential components: the Spatial Interaction Perception Module (SIPM) and the Temporal Dependency Perception Module (TDPM). The SIPM integrates a Multi-scale Hypergraph Neural Module (MHM) and a Graph Attention Module (GAM) to capture spatial interaction features across different dimensions of trajectory data. At the same time, the TDPM utilizes a deformable self-attention mechanism to identify long-range dependencies in the temporal dimension. By utilizing both SIPM and TDPM sequentially, the model captures more comprehensive spatiotemporal information. Additionally, we developed an Intention-Trajectory Fusion Module (ITFM) to account for the influence of vehicle intent on future trajectories. The ITFM combines vehicle intent probabilities with trajectory feature vectors, creating richer fused embeddings that are input into the decoder to generate future trajectory predictions, enhancing accuracy. We tested MGNformer on two publicly available vehicle trajectory datasets, NGSIM and HighD, where it achieved RMSE reductions of 5.49% and 16.67%, respectively, demonstrating a competitive performance level.

I. Introduction

In recent years, intelligent autonomous vehicles, capable of operating without human intervention, have made significant strides in reducing traffic congestion, easing driver fatigue, and enhancing road safety. However, real-world driving environments are often complex and dynamic, placing stringent demands on the safety and decision-making capabilities of these vehicles. For example, when another vehicle approaches, an autonomous vehicle must swiftly decelerate to prevent collisions. Autonomous vehicles, like human drivers, need to anticipate future traffic conditions to make informed decisions and plan their actions. Predicting the future paths of surrounding vehicles is a particularly challenging task, as it involves accounting for complex and often unpredictable interactions between road users.