Timewise Intentions and Time-Varying Distribution Network for Pedestrian Trajectory Prediction

Pedestrian trajectory prediction is crucial for intelligent surveillance, social robot navigation, and autonomous driving systems, attracting substantial research attention in recent years. Despite significant advances, accurate trajectory prediction remains challenging due to the inherent uncertainty in pedestrian intentions and the multimodal nature of human movement patterns. There remain two limitations in existing methods. First, they focus solely on predicting final goals while overlooking crucial intermediate intentions that guide pedestrian movement. Second, they utilize a static latent distribution model across all future timesteps, which fails to capture the dynamic and evolving nature of trajectory uncertainties as pedestrians move. To address these challenges, we propose a novel timewise intentions and time-varying distribution network, TITDNet, which can estimate pedestrian intentions over time while dynamically modeling trajectory uncertainties at each future timestep. Specifically, TITDNet includes two key components: an intention generator that estimates dynamic pedestrian intentions, and a variational autoencoder that captures the time-varying multimodal nature of future trajectories. A trajectory decoder then integrates historical movement patterns, predicted intentions, and learned distributions to generate accurate future trajectories. Extensive experiments on ETH, UCY, and SDD benchmark datasets demonstrate that our approach significantly outperforms the state-of-the-art methods.