Scene-aware satellite-to-ground cross-view image localization method and system

By using polar coordinate space alignment and a two-stream feature decoupling network, the geometric misalignment and feature coupling problems in cross-view image localization are solved, achieving high-precision and robust cross-view image localization, which is suitable for visual navigation of unmanned systems and disaster emergency rescue.

CN122134810APending Publication Date: 2026-06-02WUHAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
WUHAN UNIV
Filing Date
2026-04-15
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing technologies suffer from severe geometric misalignment, insufficient perception of feature aggregation saliency, and high coupling between geographic features and environmental noise in cross-view image localization, resulting in insufficient localization accuracy and robustness in complex environments.

Method used

By employing polar coordinate spatial alignment, differentiable generalized mean pooling, and a two-stream feature decoupling network, geometric misalignment is eliminated through polar coordinate transformation, salient landmark features are adaptively focused, and geographic features are decoupled from environmental noise, thereby improving positioning accuracy and robustness.

Benefits of technology

It significantly improves the cross-view positioning accuracy and robustness in complex environments, enabling high-precision feature extraction and retrieval in unmanned system visual navigation, air-space-ground collaborative perception, and disaster emergency rescue.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122134810A_ABST
    Figure CN122134810A_ABST
Patent Text Reader

Abstract

This invention discloses a satellite-ground cross-view image localization method and system based on scene awareness. First, it introduces polar coordinate mapping as an explicit geometric prior to perform deterministic spatial reparameterization of the top-view image, mitigating spatial misalignment between cross-source images. Second, it constructs a backbone network based on saliency awareness, adaptively focusing on high-response regions through a differentiable generalized mean pooling mechanism, enhancing the feature aggregation capability for key rigid geographic landmarks. Next, it designs a dual-stream visual feature decoupling architecture, jointly optimizing it with feature positive reciprocal repulsion loss and cross-view contrastive retrieval loss to forcibly decouple globally mixed features into view-invariant geospatial structure features and transient environmental noise features. Finally, it removes environmental noise and uses high-purity geographic features for accurate matching. This invention has significant practical application value in fields such as visual navigation for unmanned systems, air-space-ground collaborative perception, and disaster emergency rescue positioning.
Need to check novelty before this filing date? Find Prior Art