UAV Visual Localization System Empowered by Zero-Shot Deep Feature Matching

P. 350–354.

Viktor V. Prutyanov, Nikita L. Korolev, Aleksandr Y. Romanov

This paper presents a UAV visual localization system that leverages zero-shot deep feature matching to achieve localization in 4 degrees of freedom. The proposed system combines short-term optical flow and long-term map matching techniques, integrated through a Kalman filter for robust measurement fusion. Key contributions include the use of pre-trained deep learning models (SuperPoint, ALIKED, and LightGlue) without fine-tuning, ensuring generalization across diverse environments. Evaluated on the AdM_UAV dataset, the system achieves an average Euclidean distance of 10-14 meters, comparable to original study, while being capable for real-time operation on resource-constrained hardware like the NVIDIA Jetson Orin Nano. The system demonstrates robustness to variations in satellite maps and landscape homogeneity, though performance degrades in visually uniform areas such as fields or forests. Comparative analysis highlights the superiority of SuperPoint and ALIKED with LightGlue in balancing accuracy and computational efficiency.

Keywords: UAV, zero-shot learning, satellite imagery, visual localization, keypoint matching

In book

2025 International Russian Automation Conference (RusAutoCon)

IEEE, 2025.