Vision-Based UAV Self-Positioning in Low-Altitude Urban Environments

Area of Science:

• Computer Vision
• Robotics
• Geospatial Intelligence

Background:

• Unmanned Aerial Vehicles (UAVs) depend on satellite navigation, which can fail due to signal loss.
• Existing datasets are often unsuitable for UAV self-positioning, lacking dense sampling and real-world scenario fidelity.
• There is a need for specialized datasets to develop reliable vision-based UAV self-positioning techniques.

Purpose of the Study:

• To introduce DenseUAV, the first publicly available dataset specifically designed for UAV self-positioning.
• To provide a dataset with dense sampling of real-world, low-altitude urban UAV imagery.
• To facilitate research into robust vision-based navigation for UAVs.

Main Methods:

• Developed DenseUAV dataset with over 27,000 UAV and satellite images from 14 university campuses.
• Investigated the efficacy of Transformers over Convolutional Neural Networks (CNNs) for the self-positioning task.
• Incorporated metric learning and mutually supervised learning to improve representation and cross-modality learning.
• Introduced enhanced evaluation metrics (Recall@K, SDM@K) for retrieval and localization performance.

Main Results:

• The proposed baseline method achieved a Recall@1 score of 83.01% on the DenseUAV dataset.
• The baseline method also attained an SDM@1 score of 86.50%, demonstrating strong localization performance.
• Transformers showed superiority over CNNs for the UAV self-positioning task.

Conclusions:

• The DenseUAV dataset provides a valuable resource for advancing UAV self-positioning research.
• The proposed methods, including metric and mutually supervised learning, show promise for improving vision-based navigation.
• The developed dataset and methodologies contribute to more reliable UAV operations in GPS-denied environments.