航空学报 > 2026, Vol. 47 Issue (17): 333094-333094   doi: 10.7527/S1000-6893.2026.33094

基于深度学习的无人机视觉定位技术研究进展

李红光1, 刘泽伟2(), 李新军1, 顾颖彦3   

  1. 1.北京航空航天大学 无人系统研究院,北京 100191
    2.北京航空航天大学 电子信息工程学院,北京 100191
    3.江苏自动化研究所,南京 222061
  • 收稿日期:2025-11-17 修回日期:2025-12-25 接受日期:2026-02-27 出版日期:2026-03-17 发布日期:2026-03-16
  • 通讯作者: 刘泽伟 E-mail:13791340187@163.com
  • 基金资助:
    国家自然科学基金(62076019)

Advances in UAV visual positioning technology based on deep learning

Hongguang LI1, Zewei LIU2(), Xinjun LI1, Yingyan GU3   

  1. 1.Research Institute of Unmanned System,Beihang University,Beijing 100191,China
    2.School of Electronics and Information Engineering,Beihang University,Beijing 100191,China
    3.Jiangsu Automation Research Institute,Nanjing 222061,China
  • Received:2025-11-17 Revised:2025-12-25 Accepted:2026-02-27 Online:2026-03-17 Published:2026-03-16
  • Contact: Zewei LIU E-mail:13791340187@163.com
  • Supported by:
    National Natural Science Foundation of China(62076019)

摘要:

无人机(UAV)视觉定位是复杂环境下自主飞行的关键技术。在全球卫星导航系统(GNSS)受限或拒止条件下,传统惯性测量单元(IMU)/视觉里程计(VO)融合及人工特征匹配方法难以适应跨视角、大尺度和复杂光照场景。近年来,深度学习的发展显著提升了无人机视觉定位的精度与鲁棒性。围绕近年来基于深度学习的无人机视觉定位技术开展系统综述,梳理了相关研究的发展脉络并依据技术路线将现有方法归纳为基于模板相似度匹配、基于特征描述算子匹配和基于端到端位姿回归3大类;进一步对各类方法的关键技术进行了细化分类,系统分析了代表性方法的核心思想、技术特点、适用场景及性能表现,并比较了不同方法在定位精度、鲁棒性和实时性等方面的优势与不足。在此基础上总结了当前研究面临的地图不完备、跨视角变化、复杂环境适应性及定位可靠性等关键挑战,并对多传感器融合、运动连续性约束及智能自主定位等未来发展方向进行了展望,以期为无人机视觉定位技术研究提供参考。

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 localization is a key technology for enabling autonomous flight in complex environments. Under Global Navigation Satellite System (GNSS)-denied or GNSS-degraded conditions, conventional approaches based on Inertial Measurement Unit (IMU)/Visual Odometry (VO) fusion and handcrafted feature matching often struggle to cope with large viewpoint variations, significant scale changes, and challenging illumination conditions. Recent advances in deep learning have substantially improved the accuracy and robustness of UAV visual localization. This paper presents a comprehensive survey of recent advances in deep learning-based UAV visual localization and systematically reviews the evolution of related research. According to the underlying technical paradigms, existing methods are categorized into three groups: template similarity-based localization, feature descriptor matching-based localization, and end-to-end pose regression-based localization. Furthermore, the key techniques within each category are systematically analyzed, and representative methods are compared in terms of their fundamental principles, technical characteristics, application scenarios, and localization performance, with particular emphasis on localization accuracy, robustness, and computational efficiency. Finally, the major challenges in current research, including incomplete or mismatched maps, large viewpoint variations, adaptation to complex environments, and localization reliability, are discussed. Future research directions, such as multi-sensor fusion, motion continuity constraints, and intelligent autonomous localization, are also highlighted, with the aim of providing useful insights for the further development of UAV visual localization technologies.

Key words: deep learning, unmanned aerial vehicle, visual positioning, feature matching, pose regression

中图分类号: