Dynamic Diffusion Graph Convolutional Network With Scene-Guided Gating for Trajectory Prediction in IoT-Based Intelligent Transportation Systems

Accurate trajectory prediction is crucial in IoT-based intelligent transportation systems. This task remains challenging due to the uncertainty of agent intentions and the highly stochastic social interactions between agents. Despite the promising progress, most existing studies overlook the impact of scene heterogeneity on driving behavior, thereby limiting the predictive performance of models across diverse scenes. Moreover, the highly stochastic social interactions between agents have not been fully explored, thus weakening the reliability of predictions. To address these challenges, a novel dynamic diffusion graph convolutional network with scene-guided gating is proposed in this study. First, a pioneering scene-guided gating is introduced to capture and leverage the impact of scene heterogeneity on driving behavior. This mechanism generates scene-specific non-shared parameters to selectively filter information, thus establishing scene-guided information bottlenecks to facilitate more socially-aware trajectory prediction. Then, a dynamic diffusion graph convolutional layer is formulated to comprehensively account for the highly stochastic social interactions between agents. It models social interactions as a diffusion process on a lane graph and employs bidirectional random walks on lane attention graphs to simulate the stochastic nature of social interactions. Finally, extensive experiments on the nuScenes benchmark demonstrate the effectiveness and robustness of the proposed model.