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 interactive environments. This work proposes modular method that utilises learned model environment prediction. modularity explicitly allows unsupervised adaptation trajectory prediction models to unseen environments and new tasks by relying on unlabelled image data only. We both spatial aspects given alongside per agent motions. results more informed performance comparable state-of-the-art. highlight model's capability using benchmark pedestrian problem robot manipulation task show we can transfer predictor across these completely way. The proposed approach robust label efficient forward modelling, relaxes need full re-training