欢迎访问中国科学院大学学报,今天是

中国科学院大学学报 ›› 2026, Vol. 43 ›› Issue (6): 805-813.DOI: 10.7523/j.ucas.2025.019

• 电子信息与计算机科学 • 上一篇    下一篇

基于排名聚合的跨视角图像地理定位方向鲁棒性增强方法

盛怡宁1,2, 赵理君1, 张正1, 唐娉1   

  1. 1.中国科学院空天信息创新研究院,北京 100094
    2.中国科学院大学电子电气与通信工程学院,北京 100094
  • 收稿日期:2025-01-17 修回日期:2025-04-07 发布日期:2026-11-19
  • 基金资助:
    “十四五”民用航天技术预研项目(D040404);中国科学院空天信息创新研究院“未来之星”人才计划(2021KTYWLZX07);中国科学院青年创新促进会(2022127)

Ranking aggregation-based method for enhancing directional robustness in cross-view image geo-localization

Yining SHENG1,2, Lijun ZHAO1, Zheng ZHANG1, Ping TANG1   

  1. 1.Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China
    2.School of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing 100049,China
  • Received:2025-01-17 Revised:2025-04-07 Published:2026-11-19

摘要:

跨视角图像地理定位是一项通过多源图像实现地理位置匹配的关键技术,广泛应用于机器人导航、自动驾驶等领域。然而,现有方法多基于影像方向对齐的假设,难以适应实际场景中的方向偏差问题。为此,提出一种基于排名聚合的方向鲁棒性增强方法,通过对地面影像进行随机方向偏移,结合均值聚合和最小值聚合策略提升匹配准确性与鲁棒性。在CVUSA_360和CVACT_360数据集上的实验表明,该方法显著提升了不同模型在方向未对齐场景下的性能,尤其在R@1指标上表现突出。可视化分析进一步验证了其在增强方向鲁棒性和抑制噪声方面的优势,为复杂场景下的地理定位任务提供了高效的解决方案。

关键词: 跨视角图像地理定位, 方向未对齐, 排名聚合, 方向鲁棒性

Abstract:

Cross-view image geo-localization (CVGL) is a key technique for achieving geographic location matching through multi-source images, with broad applications in fields such as robotic navigation and autonomous driving. However, existing methods typically assume that images are directionally aligned, making them less effective in real-world scenarios with directional deviations. To address this challenge, this paper proposes a ranking aggregation-based method for enhancing directional robustness. By introducing random directional shifts to ground-level images and employing mean and minimum aggregation strategies, the proposed method improves the matching accuracy and robustness. Experiments conducted on the CVUSA_360 and CVACT_360 datasets demonstrate that this approach significantly enhances the performance of various models under directionally misaligned conditions, with particularly outstanding results in terms of the R@1 metric. Visual analysis further verifies the advantages of the method in strengthening directional robustness, suppressing noise, and improving matching accuracy, thus providing an efficient solution for geo-localization tasks in complex scenarios.

Key words: cross-view image geo-localization, directional misalignment, ranking aggregation, directional robustness

中图分类号: