Privacy Protected Multi-Domain Federated Learning for Object Detection in Autonomous Driving
- In this thesis, the application of Federated Learning in the context of object detection in autonomous driving is investigated. Specifically, the focus is on Federated Learning methods that tackle the challenge of heterogeneous client data. To select suitable Fed-erated Learning methods given a specific use-case setting, a selection model based on descriptive and explanatory models is proposed. An example use-case is applied and the derived Federated Learning methods from the proposed model are validated empirically on two autonomous driving object detection datasets. This approach demonstrates the effectiveness of the selection model on a real-world use-case and dataset. The results show that Federated Learning methods improve object detection models in autonomous driving while preserving data privacy.
This work was conducted at the Automated Driving Alliance at Robert Bosch GmbH in collaboration with the FIR e. V. at RWTH Aachen and the Institute of Imaging and Computer Vision.
