Satellite Image-based UAV Localization using Siamese Neural Network

We present a method for UAV localization using pre-existing satellite images. The use of Unmanned Aerial Vehicles (UAVs) has rapidly increased in several applications such as surveillance, search, and defense. When in GPS-denied situations, however, the onboard GPS signal may be noisy or inaccurate. The proposed method is based on a Siamese Neural Network that contains two instances of the same neural architecture and weights. Siamese Neural Network learns the similarity metric so that can recognize the same place from two raw images. Convolutional Neural Network is used as a backbone in Siamese Neural Network to overcome variation due to differences such as perspective, shadow angle, and presence of vehicles. We describe UAV localization pipeline and a dataset for training and testing our networks. Finally, the performance of the proposed method was shown in accuracy.