The paradigm of federated learning is a quickly growing and diversifying field in distributed machine learning. Through it computational resources of networked edge devices can be utilized to collectively train machine learning models under coordination of a central server. Due to local on device training with exclusively local training data and only the trained model being shared with the network no local data is directly exposed the network. Pedestrian trajectory forecasting methods model the intentions of all observed pedestrians as well as the social interactions between pedestrians to accurately predict the future trajectories of all pedestrian observed in a scene. With federated trajectory forecasting being in its infancy no public attempts have been made to federate the training of pedestrian trajectory forecasting models.

This project provides a first foray into the application of federated learning to pedestrian trajectory forecasting. A test environment consisting of a combination of the state of the art frameworks FedML in federated learning and Trajnet++ in pedestrian trajectory forecasting is developed and applied to provide initial insights into the impact of system design choices on convergence behaviour in this setting. For this an implementation of the SocialLSTM model with social grid based interaction module was federated in training over 12000 scenes from the UCY/ETH/Trajnet datasets and benchmarked through Trajnet++ evaluation tools.

Experiments in 10 federated learning system settings are evaluated in Trajnet++ final performance metrics and compared to a locally retrained baseline model of the same type provided through Trajnet++. The impact of choices in the number of communication rounds, the number of participating clients per communication round as well as the choice of adaptive or stateless federated optimization is empirically investigated.

The classicly retrained baseline outperforms all federated settings. It was found that eventhough the overall volume of training data remained the same for all experiments the data volume allocated each clients in a communication round was strong indicator for performance with higher allocated data volumes settings outperforming setting with lower allocated data volumes when the same optimization method was applied. The adaptive federated optimization algorithm FedAdam consistently outperformed Federated Averaging optimization algorithm in otherwise identical settings. Larger training datasets are necessary to further improve the performance of federated pedestrian trajectory forecasting.




  • No labels