A spatial-temporal dynamic interaction graph based transformer architecture for pedestrian trajectory prediction

Pedestrian trajectory prediction plays a critical role in intelligent transportation systems by supporting proactive operational control and infrastructure planning in complex public spaces. However, existing models struggle to capture dynamic non-local dependencies and fragmented multimodal interactions, which limits their applicability in real-world operational environments. To address these issues, this paper proposes STDIGFormer, a framework based on spatial-temporal dynamic interaction graphs and Transformer network. First, three types of dynamic interaction graphs are constructed to explicitly encode multimodal interaction patterns. Second, a Graph Transformer network is designed to combine gated memory mechanism with edge feature to enhance the structured representation of dynamic graphs. Additionally, dynamic window sparse attention mechanism is introduced to adaptively focus on key temporal segments. Furthermore, a multiscale fusion module based on co-attention and depthwise separable convolution is proposed to achieve dynamic weighting and multi-granularity complementarity across spatial-temporal features. Experimental results on ETH/UCY datasets show that STDIGFormer reduces the average and final displacement error by 17% and 31%, compared with baseline models. The ablation experiments validate the effectiveness of the method in improving prediction accuracy. Validation at Shanghai Hongqiao Transportation Hub further demonstrates the framework’s applicability to crowd density estimation, with errors below 0.1 persons/m2 in most grid cells and accurate identification of congestion hotspots. These results indicate that STDIGFormer provides an effective data-driven solution for integrating microscopic trajectory prediction with macroscopic crowd management, offering practical support for transport operations and intelligent infrastructure planning.