Learning aerial image similarity using triplet networks

Today unmanned aerial vehicles (UAV) are becoming increasingly used. For their safe travel they use GPS connection, but in recent years GPS blocking technology has improved. For this reason UAV should have a way to navigate without any external signals. One solution is to use a camera built into the UAV that is pointing to the ground. The photos taken by this camera are compared to the systems‘ known maps and by finding similarities between them determine the geographical location of the UAV. Part of this solution was explored in the research work, where the neural network VGG16 was used to discover similarities between orthographic photos. Another part of the solution was explored in another research paper, which proposed the use of a particle filter to solve the problem of UAV localization. In this paper, an algorithm will be developed based on particle filter and DenseNet-121 neural network to solve unmanned aerial vehicles localization problem. A new dataset of orthographic image triplets will be also developed to train the neural network