Federated learning-based aerial image segmentation for collision-free movement and landing

The utilization of drones has recently revolutionized remote sensing with their high spatial resolution and flexibility in capturing images. In the proposed work, we employ a swarm of drones that communicate in a wireless network. Each drone captures the image frames, and each frame is further used to locate and differentiate different objects in an image frame. The semantic segmentation of the captured images is done using deep learning algorithms. To identify the most suitable, cost-efficient, and accurate segmentation method, various state-of-the-art models, are appraised and compared based on different evaluation metrics. Resnet50 model with U-net segmentation model performs the best out of all used models by providing 91.51% pixel accuracy. Also, to give real-time predictions, we have used federated learning with the drone network. Each drone trains a local model using its accumulated data and then transfers the locally trained model to the central server that aggregates the received models, generates a global federated learning model, and transmits it in the swarm network.