Aerial-Satellite Image Matching Framework for UAV Absolute Visual Localization using Contrastive Learning

We present an Aerial-Satellite imagery matching framework for UAVs visual localization. The usage of Unmanned Aerial Vehicles (UAVs) has been continuously increasing in many applications, such as defense, agriculture, mapping, and observation. Localization plays an essential role in UAVs navigation system. The Global Positioning System (GPS) is mainly applied for localization; however, GPS interference may occur in a GPS-denied environment. One alternative that can be used is to use a pre-existing satellite image. By comparing aerial photographs with geo-tagged satellite images, global coordinates can be estimated in an intuitive way. However, due to various factors (such as weather, light conditions, and seasonal changes), it draws different visual representations between aerial and satellite imagery. To address this problem, we propose a image matching framework that exploits CNN-based Siamese Neural Network with contrastive learning method. The proposed framework firstly takes two input data; aerial image and satellite image of the area where UAV is estimated to be. By image retrieval and matching process, it predicts global coordinates of center of the aerial image. In this paper, we describe the structure of framework, data preparation, and the efficient representation learning strategy for downstream tasks. Additionally, we evaluate the performance of the proposed framework by measuring RMSE.