CSFFN: A cross-spatiotemporal feature fusion network with future interaction for vehicle trajectory prediction

Abstract

The development of autonomous driving technology relies on high-precision trajectory prediction. However, existing methods mostly focus on historical vehicle interaction characteristics while neglecting the impact of potential future interactions on dynamic behavior. In this paper, we propose a Cross-Spatiotemporal Feature Fusion Network (CSFFN) to achieve accurate prediction by fusing historical trajectory and future interaction characteristics. Specifically, we first explicitly separate the short-term dynamics and long-term trends of trajectories, and explore future interactions using graph-based attention network and dynamic channel mixing. Then, the future feature feedback is introduced to feed back the dynamic interaction weights by rough prediction of future trajectories. Furthermore, multi-scale historical trajectories and future interaction features are integrated to predict the motion trajectories of target traffic subjects. Experimental results demonstrate that our model significantly outperforms existing methods on NGSIM and High-D datasets, particularly excelling in complex scenarios by capturing long-term dependencies and potential interactions more accurately. Compared with traditional models, the key advantage of the proposed model is that it can achieve deep coupling modeling of historical-future interaction features, which provides higher safety and reliability of trajectory prediction for autonomous driving decisions.