Todor Davchev, Michael Burke, Subramanian Ramamoorthy

University of Edinburgh

arxiv link | IEEE link

abstract: 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.

Model Architecture

Conclusion

Autonomous systems operating in multi-agent environments need effective behavioural prediction models. This paper introduced a model-based solution to the problem of nonlinear, multi-agent trajectory prediction given image sequences and position information. 

The proposed architecture, RDB, was compared to existing approaches on established benchmark datasets and a newly proposed one. This work showed that both spatial and dynamic aspects of the environment are key to building effective representations in the context of multi-agent motion prediction. As a further benefit, decoupling the learning of environment specific model from the behavioural prediction component relaxes data requirements and allows for better generalisation across tasks.

Empirical evaluations along with a detailed ablation study highlighted the importance of the proposed representation modules. The modular structure builds upon a model of the static environment and the global dynamics of a scene. Results show that these models are complimentary and necessary for successful motion prediction.