摘要
无人机视角地理定位通过匹配带地理标签的卫星影像与不带标签的无人机影像实现定位与导航。现有方法侧重提取全局特征或分割区域特征,导致信息丢失。为此,提出地理定位的空间频域注意模型(SFDAM),通过频域特征学习确保图像信息完整。模型通过空间频域跨维交互(SFDI)模块提取全局空间结构和细致纹理特征,并通过显著特征提取(SFE)模块利用可学习空间注意力图,将特征分为类感知显著特征和类无关非显著特征,端到端学习多种特征,实现综合表示。结合交叉熵和三元组损失提升跨视角地理定位准确性。由University-1652和SUES-200数据集实验表明,SFDAM在无人机定位与导航任务中表现出色。
UAV-view geolocation achieves positioning and navigating by matching geotagged satellite images with non-geotagged UAV images.Existing methods focus on extracting global or segmented regional features,which may lead to information loss.To address this,a geolocation Spatial-Frequency Domain Attention Model(SFDAM)is proposed,ensuring comprehensive image representation through frequency-domain feature learning.The model uses a Spatial-Frequency Domain Interaction(SFDI)module to extract global spatial structures and texture features.It also incorporates a Salient Feature Extraction(SFE)module with a learnable spatial attention map,dividing features into class-aware salient and class-agnostic,non-salient features,learning multiple features end-to-end,and realizing integrated representations.By combining cross-entropy and triplet loss,the model enhances cross-view geolocation accuracy.Experiments on the University-1652 and SUES-200 datasets demonstrate that SFDAM performs well in UAV positioning and navigating tasks.
作者
徐海燕
赵艮平
程良伦
XU Haiyan;ZHAO Genping;CHENG Lianglun(School of Computer Science,Guangdong University of Technology,Guangzhou 510000,China)
出处
《电光与控制》
北大核心
2026年第1期58-64,共7页
Electronics Optics & Control
基金
广东省基础与应用基础研究基金(2024A1515010036)。
