Cascaded Physical-constraint Conditional Variational Auto Encoder with socially-aware diffusion for pedestrian trajectory prediction

Pedestrian trajectory prediction serves as a crucial prerequisite for various tasks such as autonomous driving and human–robot interaction. The existing methods mainly leverage deep learning-based generative models to predict future multi-modal trajectories. Nevertheless, the inherent uncertainty in pedestrian movements poses a challenge for deep generative models to generate accurate and plausible future trajectories. In this paper, we propose a two-stage trajectory prediction network termed CPSD. In the first stage, a Cascaded Physical-constraint Conditional Variational Auto Encoder is proposed. It combines Differentiable Physical Constraint Conditional Variational Auto Encoders in the cascaded form to predict the trajectory coordinates with a stepwise manner, which improves the interpretability of deep generative network and alleviates the problem of prediction error accumulation over time. In the second stage, a Socially-aware Diffusion Model is proposed to refine the initial trajectory generated in the first stage. By introducing a non-local attention mechanism and constructing a social mask, we integrate pedestrian social interactions into the diffusion model, enabling the refinement of more realistic and plausible multi-modal pedestrian trajectories. Extensive experiments conducted on the public datasets SDD and ETH/UCY demonstrate that CPSD achieves more promising pedestrian trajectories compared with other state-of-the-art trajectory prediction algorithms.