TSI-GCN: Temporal-spectral interaction enhanced graph convolutional network for pedestrian trajectory prediction

Pedestrian trajectory prediction in complex, dynamic, and crowded environments remains challenging, particularly because indirect social interactions are difficult to model and interaction estimation from raw observed trajectories is easily affected by local noise and short-term high-frequency fluctuations. For higher-order graph models, such unstable first-order affinities can be further accumulated during multi-order reasoning, thereby limiting overall prediction performance. To address this issue, we propose a Temporal-Spectral Interaction Enhanced Graph Convolutional Network (TSI-GCN) for pedestrian trajectory prediction. Specifically, an Adaptive Spectral Block (ASB) is first introduced to refine observed trajectories in the spectral domain, suppressing irrelevant high-frequency disturbances while preserving motion dynamics that are more informative for interaction estimation. Based on the refined representations, pairwise affinities between pedestrians are estimated by attention. An Interactive Convolution Block (ICB) is then used to refine the pairwise affinity responses across observed frames, improving the reliability of the first-order interaction graph before higher-order propagation. Building on this, the refined interaction graph is then propagated under a physical constraint mask for higher-order reasoning and subsequent trajectory decoding. Experimental results show that the proposed model achieves an average ADE/FDE of 0.21 m/0.34 m on ETH/UCY and 0.29 m/0.48 m on SDD, demonstrating improved prediction accuracy and more reliable social interaction modeling in complex pedestrian scenarios.