Modeling intrinsic behavioral variations with domain-adaptive spatio-temporal graph convolutional networks for pedestrian intention prediction

Accurately predicting pedestrian crossing intentions is essential for safe autonomous driving, yet generalisation across diverse environments remains challenging because pedestrian behaviour varies with cultural norms, traffic regulations, and individual decision-making. This paper proposes a behavioural domain-adaptive spatio-temporal graph convolutional network (BDAST-GCN) for pedestrian intention prediction. The framework includes two components. First, a behavioural style representation module uses an enhanced Trans-TimeGAN with dimensionality reduction and clustering to identify latent behavioural domains from dynamic pedestrian features. Second, a domain-adaptive prediction module employs a spatio-temporal graph convolutional network with distribution matching loss to align feature representations across domains while incorporating vehicle speed and contextual information. Experiments on the JAAD and PIE datasets demonstrate that BDAST-GCN effectively predicts pedestrian intentions and improves generalisation across heterogeneous traffic environments. These results indicate that the proposed framework provides a robust and adaptive solution for real-world pedestrian intention prediction in autonomous driving.