Attention-Based Road Registration for GPS-Denied UAS Navigation.

Area of Science:

• Computer Vision
• Robotics
• Geospatial Analysis

Background:

• Unmanned aerial system (UAS) navigation in GPS-denied urban environments relies on aerial image registration with road landmarks.
• Current two-stage methods (road extraction and registration) are slow and sensitive to noise.
• A need exists for more robust and efficient aerial-road registration techniques.

Purpose of the Study:

• To develop a novel, end-to-end deep learning approach for aerial-road registration.
• To improve the accuracy and robustness of UAS navigation in challenging environments.
• To demonstrate the feasibility of multitask learning for simultaneous aerial-road matching and registration.

Main Methods:

• A novel attention-based, two-branch neural network architecture with shared weights was developed.
• Input images are mapped into a common embedding space using the neural network.
• A multibranch attention module was incorporated to filter false matches and enhance accuracy.

Main Results:

• The proposed method significantly reduces mean absolute errors in rotation angle (1.24x) and translations (1.38x in x, 1.44x in y) compared to state-of-the-art approaches.
• Experimental results validate the effectiveness of the attention-based neural network.
• The approach proves the feasibility of multitask learning for aerial-road matching and registration.

Conclusions:

• The developed end-to-end deep learning model offers a more efficient and accurate solution for aerial-road registration.
• This technique enhances unmanned aerial system (UAS) geolocalization capabilities in GPS-denied settings.
• The study highlights the potential of multitask learning in computer vision applications for autonomous systems.