Modeling the Uncertainty in Pedestrians Trajectory Prediction

Pedestrian trajectory prediction attracts a great deal of attention as a fundamental research in the automatic drive system, human-robot interaction, and intelligent surveillance. This paper proposes a spatiotemporal framework based on Transformer and conditional variational autoencoder (CVAE) models to predict pedestrian trajectories while accounting for uncertainty. This paper proposes a spatiotemporal framework based on Transformer and conditional variational autoencoder (CVAE) models to predict pedestrian trajectories while accounting for uncertainty. The Transformer encoder-decoder modules capture spatial and temporal features. The CVAE compares predictions to ground truth across frames to learn motion uncertainty distributions. Experiments on two benchmark datasets demonstrate the approach reduces average displacement error by 0.04 - 0.1 m and final displacement error by 0.05 - 0.15 m compared to prior methods. The results highlight the importance of modeling uncertainty for accurate pedestrian trajectory forecasting.