Orientation-Guided Homography for Fine-Grained Cross-View Localization.

Area of Science:

• Computer Vision
• Robotics
• Geospatial Analysis

Background:

• Cross-view localization aligns ground-view images with satellite imagery for precise positioning.
• Existing homography-based methods struggle with significant orientation noise and limited refinement capabilities.
• Current methods require good initial orientation estimates, limiting their applicability.

Purpose of the Study:

• To propose a fine-grained cross-view localization method (OGH-Net) that overcomes limitations of existing approaches.
• To develop a system capable of accurate localization under challenging cross-area and unknown-orientation conditions.
• To achieve real-time performance with a favorable accuracy-efficiency trade-off.

Main Methods:

• A hybrid bird's-eye-view (BEV) transformation is used to generate BEV images with preserved geometry and expanded coverage.
• A lightweight orientation-prior module predicts a coarse yaw estimate across the full ±180° range.
• A multiscale iterative homography module refines the projection matrix through hierarchical iterations across multiple feature resolutions.

Main Results:

• OGH-Net reduces mean localization error by 11% and mean orientation error by 27% on the VIGOR dataset.
• The method further reduces mean localization error by 27% on the KITTI dataset compared to state-of-the-art.
• Real-time performance is achieved, running at up to 107 FPS on a single RTX 3090 GPU.

Conclusions:

• OGH-Net offers a significant improvement in cross-view localization accuracy and robustness.
• The proposed orientation-guided homography approach effectively handles large orientation variations.
• The method presents a practical solution for real-time applications requiring precise geospatial understanding.