Learning Structured Representations of Spatial and Interactive Dynamics for Trajectory Prediction in Crowded Scenes

Context plays a significant role in the generation of motion for dynamic agents in interactive environments. This work proposes a method that utilises a learned model of the environment for motion prediction. We show that modelling the spatial and dynamic aspects of a given environment alongside the local per agent behaviour results in more accurate and informed motion prediction. Further, we observe that this decoupling of dynamics and environment models allows for better adaptation to unseen environments, requiring that only a spatial representation of a new environment be learned. We highlight the model's prediction capability using a benchmark pedestrian prediction problem and a robot manipulation task. The proposed approach allows for robust and data efficient forward modelling, and relaxes the need for full model re-training in new environments.