The thesis presents the implementation of the currently leading method for
geolocation of unmanned aerial vehicles, in the event of a loss of the positioning system; the implementation was not publicly available. As part of the work, we created a new dataset containing pairs of images from unmanned aerial vehicles and corresponding satellite images. We focused on the use of advanced neural networks, especially the pyramid vision transformer (PVT). A key role was played by the siamese neural network for comparing patterns between the two types of images. The methodology was supported by various
optimization strategies, including the use of stratified sampling, the Hanning window, and regularization techniques. The results confirm the effectiveness of the proposed method for accurate geolocation of unmanned aerial vehicles. We conclude the work with an emphasis on key findings and the potential of the developed method.
|
