STG-KNet: A Kernel-mapping-based spatial-temporal graph convolution network for pedestrian trajectory prediction

Abstract

Predicting pedestrian trajectories in complex, dynamic, and crowded environments remains a critical challenge for autonomous driving and human-robot interaction. A pervasive challenge among existing methods is their dependence on rigid graph architectures, which hinders their capacity to model the evolving patterns of pedestrian interaction and obscures the potential features of agent-to-agent relationships. Besides, spatial and time-dependent modeling in most model is mixed, and there is a lack of structural decoupling. These issues result in fragmented reasoning and degraded performance in dense pedestrian scenarios. To address these challenges, we propose STG-KNet, a unified spatiotemporal learning framework combining sparse graph convolution with kernel-based structure modeling. STG-KNet features a dual-branch spatiotemporal encoder to decouple and independently model spatial interactions and temporal motion patterns, enhanced by biologically inspired masking strategies. It further introduces a novel Graph Convolutional Kernel Mapping (GCKM) module to convert discrete graph structures into continuous Gaussian similarity matrices, enabling adaptive edge learning and interpretable feature propagation. A Temporal Convolutional Network (TCN) decoder predicts parameters of 2D Gaussian distributions for future positions, supporting multimodal sampling. Comprehensive experiments on the ETH-UCY dataset demonstrate that STG-KNet achieves state-of-the-art accuracy (ADE=0.23, FDE=0.45), outperforming existing models while maintaining structural interpretability and high computational efficiency. In particular, the model shows exceptional generalization in dense and heterogeneous scenes, confirming the effectiveness of sparse kernel-enhanced graph reasoning in trajectory prediction.