A Cross-View Geo-Localization Algorithm Using UAV Image and Satellite Image.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Cross-view geolocation of Unmanned Aerial Vehicles (UAVs) faces challenges due to differing image sources and scene similarities.
• Existing methods struggle with interference and require robust solutions for accurate UAV positioning in satellite imagery.

Purpose of the Study:

• To develop an advanced deep learning model for improving the accuracy and robustness of UAV cross-view geolocation.
• To address the limitations of current methods by leveraging attention mechanisms and specialized post-processing techniques.

Main Methods:

• Designed a Single-Stream Pyramid Transformer (SSPT) network incorporating self-attention and cross-attention mechanisms.
• Implemented a header module for upsampling and a Gaussian weight window for improved model convergence.
• Utilized style transfer technology to simulate environmental variations for data augmentation.

Main Results:

• The SSPT-384 model achieved a Relative Distance Score (RDS) of 84.40% on the UL14 dataset, a significant improvement.
• Meter-level accuracy (MA) for 3m, 5m, and 20m increased by 12%, 12%, and 10% respectively with SSPT-384.
• The SSPT-256 model also showed performance gains, with improved RDS and MA across various thresholds.
• Demonstrated strong robustness on extended datasets including thermal infrared, nighttime, and rainy conditions.

Conclusions:

• The proposed SSPT network effectively enhances UAV cross-view geolocation accuracy by mitigating interference and refining feature interactions.
• The method exhibits excellent environmental adaptability and robustness, making it suitable for real-world applications.
• The results validate the efficacy of the attention-based transformer architecture and specialized post-processing for this task.