Intention or impulse? A vehicle trajectory forecasting model with time-series decomposition and structural context enhancement

Vehicle trajectory prediction is essential for safe autonomous driving. However, current deep learning paradigms often exhibit limitations in long-term stability, as they tend to encode motion as a monolithic feature, failing to disentangle enduring driving intentions from transient reactive fluctuations. This conflation of distinct temporal dynamics compromises the accuracy. To address this challenge, we propose the Decomposition-Intention-based Multi-scale Enhancement Network (DIME-Net), which explicitly decouples these multi-scale temporal dynamics. Specifically, the architecture incorporates a driving intention perception module that leverages the Discrete Fourier Transform (DFT) to decompose motion signals into trend and periodic components, effectively isolating multi-scale intentions. This internal representation is processed in parallel by two modules. The environmental perception module contextualizes it by applying graph centrality metrics to dynamic vehicle graphs to identify key surrounding vehicles. Concurrently, the social interaction and enhancement module augments it by refining spatiotemporal dependencies through polar pooling and gated causal convolutions. These multi-source representations are subsequently fused through an adaptive gating mechanism. Extensive experiments on the NGSIM and HighD datasets demonstrate that DIME-Net achieves state-of-the-art performance and significantly reduces prediction errors. These results underscore the efficacy of explicitly modeling decoupled temporal dynamics for robust trajectory forecasting.