Robust and Cost-Effective Vision-Based Indoor UAV Localization with RWA-YOLO

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Indoor localization for unmanned aerial vehicles (UAVs) is difficult in GPS-denied areas, especially for small objects and low-light conditions.
• Existing methods often struggle with accuracy and robustness in challenging indoor environments.

Purpose of the Study:

• To develop a robust, vision-based localization framework for UAVs in GPS-denied indoor environments.
• To enhance small-object detection and multi-scale feature representation for improved localization accuracy.

Main Methods:

• Proposed Robust Wavelet-Aware YOLO (RWA-YOLO) detection framework.
• Integrated a wavelet-aware attention fusion module and dual multi-path aggregation.
• Utilized UAV-mounted LEDs for low-light visual perception.
• Employed multi-view geometric triangulation for 3D position estimation without external beacons.

Main Results:

• Achieved centimeter-level localization accuracy (RMSE: 9.9 mm, 95th percentile: 13.5 mm).
• Demonstrated robust performance in low-light and dynamic flight conditions.
• Real-time performance with an update rate of approximately 25 FPS.
• Outperformed state-of-the-art vision-based methods and comparable to hybrid systems.

Conclusions:

• RWA-YOLO offers effective, robust, and real-time indoor UAV navigation.
• The system provides centimeter-level accuracy, suitable for real-time control loops.
• Validated for static and dynamic flight conditions in real indoor environments.