Oriented aircraft object detector using Scaled YOLOv4 on very high resolution satellite and synthetic datasets
Benjamin Palmaerts, Benjamin Beaumont, Graur Dimitri, Swinnen Gérard, Eric J. Hallot · 2023
We propose a deep learning solution in the frame of the management of airport ground flows. The developed tool aims to automatically detect aircrafts on satellite imagery at very high resolution (<1m). It relies on the Scaled YOLOv4 deep learning model which has proven to be a good object detector. Trained on open-source annotated datasets, our model shows very good performance to detect common commercial airplanes on Pléiades images, but misses small planes or other types of aircrafts. We indicate several possible improvements, thanks to a large dataset of Pléiades images that we annotated ourselves and thanks to the production of synthetic images combining Pléiades data with artificial superimposed aircrafts.