Domain-Adaptive Graph Transformer for Multi-Vehicles Trajectory Prediction

Abstract: The rapid development of intelligent transportation systems (ITS) and advanced sensor technologies has made multi-agent trajectory prediction a critical component for various traffic applications, such as autonomous driving, traffic management, and collision avoidance. However, a major challenge in trajectory prediction is that models trained on specific traffic scenarios often fail to generalize effectively to unseen environments due to domain discrepancies in spatial configurations and traffic dynamics. To address this challenge, we develop a novel Graph-Based Transformer Network that uniquely integrates Domain Adaptation techniques to significantly improve the generalization of trajectory prediction models across diverse traffic domains. Our approach begins by constructing graph-based embeddings that capture the spatial relationships between vehicles, which are then processed through a Transformer Network designed to model the temporal dependencies in vehicle movements. The self-attention mechanism within the Transformer is specifically optimized to highlight critical temporal patterns, thus enhancing the accuracy of long-term trajectory predictions. A key innovation of our model lies in the incorporation of a domain adversarial training strategy, which aligns spatial-temporal features between the source and target traffic domains. This alignment substantially mitigates domain shift, ensuring consistently high performance in previously unseen traffic scenarios while significantly reducing prediction errors across diverse traffic environments.

The proposed method is rigorously evaluated in two essential cross-domain traffic settings: cross-city and cross-period prediction scenarios. For the cross-city scenario, we assess the model’s generalization capabilities on two major highway datasets, NGSIM-I80 and NGSIM-US101, which differ significantly in spatial layout and traffic dynamics, ranging from free flow to heavy congestion. In the cross-period scenario, we test the model's adaptability to varying traffic conditions, capturing transitions between low-congestion (free flow) and high-congestion periods on the same freeway. Experimental results reveal that our Graph-Based Transformer Network, augmented with domain adaptation, consistently outperforms state-of-the-art baselines such as LSTM, CNN-LSTM, and standard Transformer models. In the cross-city scenario, our model, combined with the domain adaptation strategy, demonstrated RMSE improvements of 20.81% and 21.58% over the Graph Embedded Transformer baseline in the source and target domains, respectively. Across different time periods, we observed further RMSE improvements of 17.94% and 19.64% in the source and target domains, respectively, showcasing the model's enhanced ability to generalize across varying traffic conditions.

In conclusion, the proposed Graph-Based Transformer with Domain Adaptation provides a robust solution for vehicle trajectory prediction, demonstrating strong generalization across diverse traffic conditions and spatial layouts, making it highly suitable for real-world intelligent transportation systems.