摘要
跨视角图像地理定位(Cross-view image geo-localization)是将待查询图像与已知位置不同拍摄角度的图像进行匹配,以确定图像拍摄位置的技术,在公共安全、网络空间治理等领域具有广泛应用前景.本文将现有跨视角图像地理定位方法分为传统的基于人工设计特征的方法和当前流行的基于深度神经网络的方法,并根据特征提取前的视角对齐方式对后者作进一步分类.然后,分析总结了不同类别定位方法的优缺点,为综合利用提供依据.分析结果表明,视角对齐能够有效弥补不同视角之间的巨大差异,提高图像检索的效率和精度,因此设计可解释性强且失真小的视角对齐方法是跨视角图像定位技术研究重点突破的问题之一.
Cross-view image geo-localization is a technique that matches the queried street-view image with referenced satellite image datasets,to determine the geo-location of camera,and can be applied in the field of public security operations,military intelligence and cyberspace governance.Existing cross-view image geo-localization methods can be classified into traditional methods based on hand-crafted features and the currently popular methods based on deep neural networks.Furthermore,the latter can be further classified in terms of the view alignment methods used before feature extraction.Then,the advantages and disadvantages of different categories of localization methods are analyzed and summarized to provide a basis for their combined utilization.The analysis results show that view alignment can effectively make up for the huge differences between different views and improve the efficiency and accuracy of image retrieval.Therefore,designing view alignment methods with high interpretability and low distortion is one of the key breakthrough issues in cross-view image geo-localization.
作者
刘楝
李江杉
许经乾
杨春芳
秦雨林
刘琰
LIU Lian;LI Jiangshan;XU Jingqian;YANG Chunfang;QIN Yulin;LIU Yan(Political and Legal Committee of the Central Military Commission,Beijing 100000,China;Zhengzhou University,Zhengzhou 450002,China;Information Engineering University,Zhengzhou 450001,China)
关键词
跨视角图像地理定位
图像地理定位
计算机视觉
cross-view image geo-localization
image-based localization
computational vision
