A dual-graph evidential transformer framework for highway vehicle trajectory prediction

Accurate vehicle trajectory prediction in highway scenarios is essential for improving traffic safety and efficiency. However, drone footage allows for wide-area monitoring but is subject to localization noise, intermittent occlusions, and scenario changes. Consequently, it is essential to forecast both a trajectory and its reliability. This study introduces GNIG-Net, a graph-based neural architecture that models fine-grained vehicle–vehicle and vehicle–lane dynamics captured from overhead drones. GNIG-Net couples dual heterogeneous graphs with a spatial-temporal transformer to learn rich interaction patterns. Additionally, it incorporates a Normal-Inverse-Gamma evidential layer that produces closed-form, separate estimates of aleatoric (data-driven) and epistemic (model-driven) uncertainty. The resulting confidence intervals can be fed directly into safety-critical functions such as collision avoidance or variable-speed control. Experiments conducted on three publicly available datasets (highD, exiD, A43) demonstrate that GNIG-Net outperforms seven state-of-the-art baselines. Specifically, on the highD dataset, the proposed model reduces the Average Displacement Error (ADE) by 18% and the Final Displacement Error (FDE) by 13% compared to the strongest Transformer-based competitor. These results highlight GNIG-Net's potential to enhance real-world traffic management and autonomous driving systems.