UAV Geo-Localization Dataset and Method Based on Cross-View Matching.

Area of Science:

• Computer Vision
• Robotics
• Geospatial Intelligence

Background:

• Drones rely on Global Navigation Satellite Systems (GNSS) for stable flight.
• GNSS signals are unreliable in complex environments, causing flight instability.
• Cross-view machine learning offers potential for robust drone localization.

Purpose of the Study:

• To introduce a new dataset (VDUAV) for cross-view drone localization.
• To develop a novel network architecture (VRLM) for improved geolocation.
• To address limitations of existing datasets and localization methods.

Main Methods:

• Creation of the VDUAV dataset using a digital twin platform and virtual-real mapping.
• Development of the Virtual Reality Localization Method (VRLM) with FocalNet backbone.
• Feature extraction from drone and satellite images using separate branches and multi-scale feature fusion (SCFF module).

Main Results:

• The VDUAV dataset significantly reduces dataset production costs.
• The VRLM model achieved 83.35% accuracy (MA@20) and 74.13% precision (RDS) on the VDUAV dataset.
• VRLM demonstrated superior performance compared to the FPI baseline.

Conclusions:

• The VDUAV dataset and VRLM network provide a robust solution for cross-view drone localization.
• The proposed methods enhance drone navigation stability in GNSS-denied environments.
• This work opens new avenues for cross-view geolocation research and applications.