LANDMARKS DETECTION USING A NEURAL NETWORK DURING A UAV FLIGHT ALONG A GIVEN ROUTE
E. S. Nejevenko, G. I. Gromilin, V. P. Kosykh
Institute of Automation and Electrometry, Siberian Branch, Russian Academy of Sciences, Novosibirsk, Russia
Keywords: UAV video camera, landmark, satellite image, detection, YOLOv12, augmentation, training set
Abstract
Cross-view geolocalization is the process of establishing correspondence between images of ground objects obtained by surveillance equipment deployed on different platforms, provided that the images generated by the equipment on one platform are geodetically referenced. A cross-view geolocalization method is proposed based on the detection of landmarks e field of view of an unmanned aerial vehicle (UAV) video camera using the YOLOv12 neural network, trained on georeferenced satellite images containing these landmarks. The training set is formed by augmenting the images of the specified landmarks. The results of experiments obtained using a sample from the SUES-200 dataset containing 20 satellite images of various landmarks and 20×50 images of these landmarks obtained by a UAV onboard camera from an altitude of 300 m are presented.
|
