Federated meta learning for visual navigation in GPS-denied urban airspace

In this paper, we have proposed a novel FLVO framework which can improve pose estimation accuracy in terms of translational and rotational RMSE drift while reducing security and privacy risks. It also enables fast adaptation to new conditions thanks to the aggregation process of the local agents which operate in different environments.

In addition, we have shown that it is possible to transfer an end-to-end visual odometry agent that is trained by using ground vehicle dataset (i.e. KITTI dataset) to an aerial vehicle pose estimation problem for low-altitude and low-speed operating conditions.

Dataset size is an important topic that should be considered in both AI-based end-to-end visual odometry applications and federated learning approaches. Although it is demonstrated that federated learning could be applied for visual odometry applications to aggregate the agents that are trained in different environments, more data should be collected to improve the translational and rotational pose estimation performance of the aggregated agents.

In our future work, we will evaluate cyber-attack detection performance of the proposed FLVO framework by utilizing multiple learning loops. In addition, dataset size will be expanded by utilizing real flight tests to increase the realm of the training data and to improve the robustness of the proposed federated learning based end-to-end visual odometry algorithm.