【作者】 宋涛；

【导师】 褚金奎；

【作者基本信息】 大连理工大学 ， 机械工程， 2025， 硕士

【摘要】 随着无人机在近地遥感、装备运输及灾害救援等领域的广泛应用,无人机在复杂环境下的自主定位能力成为制约其任务可靠性的核心问题。传统依赖全球导航卫星系统(Global Navigation Satellite System,GNSS)的定位方法在信号拒止场景中易失效,视觉同步定位与建图(Visual Simultaneous Localization and Mapping,V-SLAM)技术受限于低空稀疏特征环境下的图像匹配精度与实时性,难以满足实际需求。为此,本文提出一种基于卫星模板匹配的视觉地理定位方法,旨在解决GNSS拒止环境下无人机定位精度与效率的优化问题。首先,本文针对卫星影像与无人机实时影像特征差异,设计了基于Transformer架构的由粗到精匹配网络(Coarse-to-Fine Matching Network,C2FM-NET),利用注意力机制增强异源图像特征匹配的能力。基于C2FM-NET网络,本文设计了分层框架地理定位系统,在图像输入阶段,使用质量处理和清晰度处理降低图像之间特征差异;在初始位置检索阶段,针对暴力搜索算法复杂度高的问题,提出基于层次化检索树的图像检索方案降低计算复杂度;在位置追踪过程,开发了一种基于影像序列的追踪算法,通过卡尔曼滤波提升自主定位精度。其次,为验证系统有效性,本研究搭建了软硬件协同的四旋翼无人机实验平台,硬件端集成无人机平台、地面站以及通信模块,软件端基于机器人操作系统(Robot Operating System,ROS)设计多线程定位框架,实现图像采集、特征匹配与位置航向解算的实时处理。最后,利用四旋翼无人机实验平台,本文进行了无人机轨迹定位仿真与实验对比,测试不同高度和场景下本文定位方法性能,仿真结果验证了本文地理定位方法中位置检索算法和位置追踪算法的有效性。实验结果表明,在150 m高度下,本文检索方法相比暴力搜索耗时从13.84 s缩短至6.48 s,位置均方根误差为12.4 m,航向角度均方根误差为10.4°,位置追踪更新时间为0.128 s。仿真和实验证明了该方法能够满足无人机在低空稀疏特征且GNSS拒止环境下的实时地理定位需求。更多还原

【Abstract】 With the widespread application of unmanned aerial vehicles(UAVs)in near-earth remote sensing,equipment transportation,and disaster relief,their autonomous positioning capability in complex environments has become a critical factor affecting mission reliability.Traditional Global Navigation Satellite System(GNSS)-dependent positioning methods are prone to failure in signal-denied scenarios,while Visual Simultaneous Localization and Mapping(V-SLAM)technology struggles to meet practical requirements due to limitations in image matching accuracy and real-time performance in low-altitude sparse-feature environments.To address this,this paper proposes a visual geolocation method based on satellite template matching,aiming to optimize UAV positioning accuracy and efficiency in GNSS-denied environments.First,considering the feature differences between satellite imagery and real-time UAV imagery,we design a Transformer-based Coarse-to-Fine Matching Network(C2FM-NET)that enhances cross-domain image feature matching through attention mechanisms.Building upon C2FM-NET,we develop a hierarchical geolocation framework:During image input,quality enhancement and sharpness processing are applied to reduce inter-image feature discrepancies;For initial position retrieval,we propose a hierarchical search tree-based image retrieval solution to reduce computational complexity compared to brute-force search algorithms;In position tracking,we implement a sequence-based tracking algorithm that improves autonomous positioning accuracy through Kalman filtering.Second,to validate system effectiveness,we establish a quadrotor UAV experimental platform with hardware-software co-design.The hardware integrates UAV platforms,ground stations,and communication modules,while the software implements a multi-threaded positioning framework based on Robot Operating System(ROS)for real-time image acquisition,feature matching,and position/orientation calculation.Finally,using the quadrotor UAV experimental platform,the UAV trajectory positioning simulation and experimental comparison are carried out to test the performance of the proposed localization method at different altitudes and scenarios,and the simulation results verify the effectiveness of the position retrieval algorithm and position tracking algorithm in the geolocation method in this paper.Experimental results show that at an altitude of 150 m,the retrieval time of the proposed method is shortened from 13.84 s to 6.48 s,the root mean square error of position is 12.4 m,the root mean square error of heading angle is 10.4°,and the update time of position tracking is 0.128 s.Simulations and experiments show that the proposed method can meet the real-time geo-positioning requirements of UAVs in low-altitude sparse features and GNSS rejection environments.更多还原