Multimodal geolocation method and system based on three-dimensional conditional cue learning

By employing a 3D conditional cueing learning method, combined with a visual-language model and a hybrid expert network, the problem of insufficient information fusion in cross-view geolocation was solved. This enabled high-precision UAV and satellite image positioning, reduced data annotation costs, and improved positioning accuracy and robustness.

CN122244168APending Publication Date: 2026-06-19HUNAN NORMAL UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HUNAN NORMAL UNIVERSITY
Filing Date
2026-05-25
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

In cross-view geolocation, insufficient information fusion and weak perception of ground feature distribution and modal differences make it difficult for existing methods to adapt to the huge viewpoint differences between UAVs and satellite images. Furthermore, the lack of semantic text annotation limits the application of multimodal methods.

Method used

A multimodal geolocation method based on 3D conditional cue learning is adopted. It uses a pre-trained visual-language model to generate semantic description text, combines a hybrid expert network and a 3D conditional cue mechanism to dynamically fuse visual and text features, and achieves efficient multimodal fusion through attention mechanism and multi-scale feature refinement.

Benefits of technology

Achieving high-precision cross-view geolocation in the absence of labeled data enhances the discriminative power and robustness of features, significantly improving positioning accuracy and fusion efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122244168A_ABST
    Figure CN122244168A_ABST
Patent Text Reader

Abstract

This invention discloses a multimodal geolocation method and system based on 3D conditional cue learning. The method includes extracting semantic descriptive text from UAV images; extracting visual features from the UAV images and textual features from the semantic descriptive text; fusing the visual and textual features of the UAV images using a hybrid expert fusion model to obtain multimodal fused features; constructing 3D conditional cue based on the UAV images, consisting of 2D spatial cues and a unified conditional vector, injecting it into the multimodal fused features through an attention mechanism, and then performing similarity matching with a pre-built feature library. The geographical location corresponding to the most matching multimodal fused feature in the feature library is output as the geolocation result. This invention aims to solve the problems of insufficient information fusion and weak perception of ground feature distribution and modal differences in cross-platform geolocation, achieving accurate geolocation across different perspectives.
Need to check novelty before this filing date? Find Prior Art