关键词: 深度学习, 无人机, 视觉定位, 特征匹配, 位姿回归

Abstract: Unmanned aerial vehicle (UAV) visual positioning, as a key technology for ensuring autonomous flight and mission execu-tion, holds significant application value in complex environments with limited or denied global satellite navigation system (GNSS) signals. Traditional fusion methods based on inertial measurement units (IMUs) and visual odometry (VO), as well as image matching and geometric positioning techniques relying on artificial features, can provide effective support in certain scenarios but generally face issues of insufficient robustness and accuracy degradation under cross-view, large-scale varia-tions, and complex lighting conditions. In recent years, with the advancement of deep learning, learning-based UAV visual positioning methods have achieved remarkable progress. This paper systematically reviews the latest research developments in deep learning-based UAV visual positioning, categorizing them into three types. For each approach, the core concepts, advantages, and limitations are analyzed, along with a summary of typical research outcomes and their performance metrics. Finally, the issues of positioning reliability under incomplete or mismatched maps, the influence of strong viewpoints and complex environmental changes on the stability of visual positioning, as well as the crucial role of motion continuity and flight constraints in denied environments were discussed. These findings provide a reference for the future intelligent auton-omous navigation of unmanned aerial vehicles.

Key words: Deep learning, Unmanned aerial vehicle, Visual positioning, Feature matching, Pose regression