AGEN: Adaptive Error Control-Driven Cross-View Geo-Localization Under Extreme Weather Conditions

Area of Science:

• Computer Vision
• Robotics
• Geographic Information Systems

Background:

• Cross-view geo-localization matches geographic images from different viewpoints (e.g., drone, satellite) without GPS.
• Existing methods struggle with performance degradation under extreme weather conditions.
• Drone-based localization and navigation demand robust geo-localization solutions.

Purpose of the Study:

• To propose AGEN, a robust end-to-end image retrieval framework for cross-view geo-localization under extreme weather.
• To enhance accuracy and reliability in challenging environmental conditions.

Main Methods:

• Integration of the DINOv2 network for global feature extraction with the Local Pattern Network (LPN) for detailed classification features.
• Introduction of an Adaptive Error Control (AEC) module utilizing fuzzy control to dynamically optimize the loss function by adjusting loss weights.
• Development of a robust end-to-end image retrieval framework (AGEN).

Main Results:

• AGEN achieved 91.71% Recall@1 accuracy on the University160k-WX dataset under extreme weather.
• State-of-the-art Recall@1 accuracy was reached on University-1652 and SUES-200 datasets for drone-view target localization and navigation.
• AGEN attained 95.43% Recall@1 on the University-1652 dataset for drone-view target localization.

Conclusions:

• The proposed AGEN framework demonstrates superior robustness and accuracy for cross-view geo-localization in extreme weather conditions.
• AGEN effectively handles complex and challenging scenarios, outperforming existing models.
• The integration of DINOv2, LPN, and AEC contributes to significant advancements in drone-based localization and navigation.