Improving the Localization Accuracy in Internet of Drones Networks

Abstract

The Internet of Drones (IoD) has emerged as a critical research focus across academia and industry due to its diverse applications ranging from civilian services to military operations. In such networks, accurate localization of Unmanned Aerial Vehicles (UAVs) is essential for effective realtime communication and navigation. While Global Navigation Satellite Systems (GNSS) are commonly used for localization, their accuracy can be severely degraded in complex or obstructed environments. To tackle this challenge, this paper presents a robust cross-view geo-localization framework leveraging Convolutional Neural Networks (CNNs) to match UAV and satellite images. The model employs ResNet-101 as a backbone to extract deep spatial-semantic features and is trained on the University-1652 dataset using synthetic UAV imagery. Cosine similarity is utilized to measure directional closeness between feature embeddings, enhancing retrieval robustness. Experimental results confirm the effectiveness of the proposed framework, achieving a Recall@1 of 78.23%, a Recall@5 of 86.35%, a Recall@10 of 92.11%, and an Average Precision (AP) of 83.65%. These outcomes highlight the robustness of the model in cross-view image matching, positioning it as a practical solution for UAV localization in IoD scenarios where GNSS signals are unreliable.