Secure Vision-Based Navigation for Drones in GPS-Denied Environments Using Machine Learning
Durbovnic, Croatia
Unmanned Aerial Vehicles (UAVs) have become indispensable across domains such as surveillance, disaster management, and industrial inspection. However, their heavy reliance on Global Positioning System (GPS) signals exposes them to spoofing, jamming, and denial-of-service attacks, creating significant risks in critical operations. To address this limitation, we propose a secure vision-based navigation framework that integrates Convolutional Neural Networks (CNNs), Simultaneous Localization and Mapping (SLAM), and efficient path-planning algorithms. A lightweight UNet architecture with a ResNet50 backbone was trained on aerial datasets, achieving a Dice coefficient of 0.9155 and a mean Intersection over Union (mIoU) of 0.8658, enabling robust segmentation of roads, buildings, terrain, and obstacles. SLAM, powered by SuperPoint feature detection and SuperGlue matching, demonstrated reduced drift and improved trajectory consistency compared to classical ORBbased methods. Path-planning experiments further showed that A* consistently outperformed Dijkstra in real-time navigation scenarios. The integration of CNN-driven perception with SLAMbased localization provides reliable cost-fused maps for autonomous decision-making. This framework enhances confidentiality, integrity, and availability in UAV navigation, offering a scalable GPS-independent solution for secure deployment in contested and GPS-denied environments.