Learning aerial image similarity using triplet networks

The main way to determine the location of an aircraft is to use a GPS signal. However, this method works inaccurately when communication with satellites is lost or contaminated with interference, so it is important to explore alternative ways to locate an aircraft. One such method is explored in this research paper, which is to use a particle filter algorithm to search for the location of the aircraft based on the orthographic images taken by the aircraft. This algorithm must be able to determine whether two orthographic photographs are similar or not. For this task, a triplet neural network with ResNet base layers is used. Its accuracy is compared to the accuracy of the neural network with VGG-16 base layers. A dataset of Vilnius orthographic images is created to train the neural network and to perform experiments. Drone flight simulations are used to determine which parameter values provide the highest accuracy of the drone location recognition algorithm