Deep Learning Based Autonomous Navigation in GPS Denied Environments Using Landmark Recognition

Abstract

Reliable navigation is a fundamental requirement for unmanned aerial vehicles (UAVs), especially in long-range missions where Global Positioning System (GPS) signals may be unavailable, unreliable, or intentionally denied. This thesis proposes a deep learning-based GPS-denied navigation system designed to enable robust long-range localization for UAVs using OpenStreetMaps to achieve robust and reliable localization. The proposed approach first performs coarse localization by matching UAV imagery with OpenStreetMap-aligned representations to identify a candidate geographic region. A second fine localization stage refines the estimated position using a convolutional neural network. The results show that integrating deep learning with OpenStreetMap-based priors provides a scalable and effective solution for GPS-denied UAV navigation, offering a practical step toward reliable autonomous flight in complex and unstructured environments