关键词: 交通运输工程；行人轨迹预测；多模态预测；注意力机制；多阶段学习

Abstract: In complex and dynamic traffic environments, accurately predicting pedestrian trajectories is crucial for fields such as autonomous driving, service robots, and smart cities. However, the high uncertainty of trajectory distribution, especially the spatial distribution differences between short-term and long-term predictions, poses challenges for modeling. To address this, a three-stage pedestrian trajectory prediction framework incorporating the idea of ITransformer, ITER (inverse trajectory transformer), was proposed. Through a staged strategy, the motion patterns at different time scales were captured in a hierarchical manner. Firstly, through short-term prediction, local dynamic patterns were learned. Secondly, the potential distribution of endpoints was modeled to grasp the global trend of long-term trajectories. Finally, fusing the information from the first two stages, the fine modeling of the complete future trajectory was achieved. Experimental results show that, in five scenes of the ETH/UCY dataset, ITER reduces the average displacement error by 13.04% and the final displacement error by 20.51% compared to the Agentformer model. In the SDD dataset, the proposed method reduces the average displacement error by 8.28% and the final displacement error by 8.02% compared to the Y-Net model. The research indicates that the proposed model achieves significant performance improvement, which can generate accurate and stable future trajectories, achieving the precise prediction for pedestrian trajectories.

Key words: traffic and transportation engineering; pedestrian trajectory prediction; multimodal prediction; attention mechanism; multi-stage learning