Transferable multi-level spatial-temporal graph neural network for adaptive multi-agent trajectory prediction

Abstract

Accurately predicting future multi-agent trajectories at intersections is crucial yet challenging due to complex and dynamic traffic environments. Existing methods struggle with cross-domain trajectory prediction owing to: 1) there are significant differences in spatiotemporal features between domains, which leads to insufficient modeling of trajectory temporal sequence dynamics during cross-domain spatiotemporal alignment; and 2) the strong heterogeneity of behavioral patterns within different datasets causes significant domain shifts, resulting in a notable performance decline when the model is transferred across datasets. To address the aforementioned challenges, this paper proposes a transferable multi-level spatial-temporal graph neural network (T-MLSTG). Based on maximum mean discrepancy theory, we design a windowed mean gradient discrepancy (WMGD) metric that incorporates mean and gradient information of temporal features to better capture cross-domain distribution differences. Furthermore, a multi-level spatial-temporal graph network (MLSTG) is designed with a two-level architecture. The first level encodes historical spatiotemporal features independently, while the second level integrates spatiotemporal features and employs a channel attention mechanism to enhance feature discrimination. The performance of T-MLSTG was evaluated on the inD and INTERACTION datasets. Compared to the baseline model, the cross-domain trajectory prediction results demonstrate a reduction in root mean square error (RMSE) of 0.812. In cross-dataset trajectory prediction evaluation, the mean error was reduced by 27.8%, demonstrating the method’s effectiveness and generalization capability.