SIAT: Pedestrian trajectory prediction via social interaction-aware transformer

Abstract

Abstract Pedestrian trajectory prediction is crucial for mitigating collision risks in intelligent transportation and surveillance systems. Despite recent advances, accurately capturing and modeling complex social interactions among pedestrians remains a challenge. This paper introduces the Social Interaction-Aware Transformer (SIAT), a novel approach that leverages a Transformer encoder to process pedestrian embedding features and a Graph Convolutional Network (GCN) to construct a social graph for extracting spatial interaction features. The future pedestrian trajectory is predicted using a Transformer decoder that integrates both pedestrian embeddings and social graph features. Extensive experiments on the ETH/UCY and Stanford Drone datasets demonstrate that SIAT significantly outperforms state-of-the-art methods in terms of accuracy and robustness, particularly in densely populated environments. SIAT’s contributions include improved precision through temporal and spatial processing, deep contextual understanding of pedestrian dynamics, and robustness across various settings. The novel model framework establishes a new benchmark for mixed models in trajectory prediction.