An Online Federated Learning Framework via Over-the-Air Computing for UAV Object Detection

Thanks to their mobility and flexibility, unmanned aerial vehicles (UAVs) could capture images from various angles, serving as training samples to enhance the accuracy of object detection tasks. As a distributed learning framework, federated learning (FL) can be applied to UAV networks to enable UAVs to collaboratively learn a shared model without transmitting raw data. However, the limited resources and mobility of UAVs pose stringent challenges for implementing FL, such as data drift among different UAVs and model aggregation under dynamic conditions. To address these issues, in this paper, we propose an online FL-based framework for UAV object detection tasks with misaligned over-the-air computation (AirComp) techniques. Specifically, we consider a multi-UAV scenario where each edge UAV performs local training while a central UAV is deployed for coordination and model aggregation. Considering the deviation of different viewpoints or data distribution of UAVs, we integrate the retraining-inference process into FL that retrains online collected samples to mitigate the impact of data drift. To further improve FL efficiency, we introduce the AirComp to perform fast uplink model aggregation and use maximum likelihood estimation to obtain estimated sequences to overcome the problems caused by dynamic channel conditions. Simulation results show that the proposed method improves 30% detection accuracy and reduces 52% convergence time compared to the baseline methods in UAV swarms.