Visual Localization Using Aerial-Ground Models for Precise Positioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Unmanned aerial vehicle/satellite base map-based visual localization methods suffer from low localization success rate and accuracy, particularly in large-scale scenarios.
Innovation Solution
A method that determines an initial pose based on an aerial model using skyline and semantic information of building lines and surfaces, followed by fine localization on a ground model to enhance accuracy and success rate, utilizing a server to process images and provide virtual-object description information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If unmanned aerial vehicle/satellite base map-based visual localization method is used, then large-scale scenario localization is enabled, but localization success rate and accuracy are low
Solution Approach 1:
The patent divides the localization system into two segments: aerial model-based localization for large-scale coverage and ground model-based localization for high-precision positioning. The server first performs aerial model matching to obtain initial pose, then performs ground model matching to refine the pose, achieving both wide coverage and high accuracy through segmented processing
Solution Approach 2:
The patent implements nested localization by embedding ground model-based fine localization within aerial model-based coarse localization. The aerial model provides initial positioning at larger scale, while the ground model nested within it provides refined positioning at smaller scale, allowing the system to achieve both large-scale coverage and high-precision localization
2Productivity
If aerial model-based visual localization is performed, then fast and efficient coarse localization is achieved, but localization accuracy is insufficient
Solution Approach 1:
The patent applies preliminary action by first performing aerial model-based localization to obtain initial pose information quickly, then using this preliminary result as the starting point for ground model-based refinement. This two-stage approach ensures fast initial localization followed by accuracy improvement
Solution Approach 2:
The patent changes the scale parameter from large-scale aerial model to small-scale ground model between the two localization stages. The aerial model operates at a broader spatial scale for speed, while the ground model operates at a finer spatial scale for accuracy, achieving both requirements through parameter transformation
Data Source
Figure 1(a)~1(b)
Figure 2(a)~3
Figure 4A~4B
AI summary
A visual localization method and apparatus are provided. The method includes: obtaining a captured first image, determining a first pose based on the first image and an aerial model, determining whether a ground model corresponding to the first pose exists in an aerial-ground model, andwhen the ground model corresponding to the first pose exists, determining a second pose based on the ground model. The aerial-ground model includes the aerial model and the ground model mapped to the aerial model, a coordinate system of the ground model is the same as a coordinate system of the aerial model, and localization accuracy of the second pose is higher than localization accuracy of the first pose. Performing ground model-based fine visual localization can improve accuracy and a success rate of localization.