A Fourier attention network for vehicle trajectory prediction with multi-dimensional information fusion

Vehicle trajectory forecasting plays a critical role in decision-making and motion planning for autonomous driving systems. However, accurately predicting future trajectories remains challenging because of complex spatiotemporal interactions among vehicles and the uncertainty of future driving behaviors in dynamic traffic environments. To address these challenges, this study proposes a Fourier attention-based Fastformer generative adversarial network, termed FF-GAN, for multi-step vehicle trajectory prediction. The proposed framework integrates time-domain and frequency-domain feature representations to enhance spatiotemporal feature learning. Specifically, a Fourier spatiotemporal attention module is developed to capture fine-grained temporal and spatial dependencies in the frequency domain. The resulting attention operators are further incorporated into the Fastformer architecture to improve global interaction modeling. In addition, a generative adversarial learning framework is employed to generate multiple plausible candidate trajectories and represent the uncertainty of future vehicle motion. Extensive experiments are conducted on the NGSIM and highD datasets. FF-GAN achieves the lowest RMSE across all evaluated prediction horizons on NGSIM, with an average RMSE of 1.36 m, representing a 6.2% reduction compared with the second-best method. On highD, FF-GAN obtains the lowest average RMSE of 0.68 m, reducing the average error by 2.9% compared with the second-best method. These results confirm the effectiveness of FF-GAN for vehicle trajectory forecasting under different highway traffic conditions. The code is publicly available at: https://github.com/gyhhq/Fourier-trajectory-prediction .