Single-Image Camera Pose Estimation from Overhead Feature Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous and semi-autonomous vehicles face challenges in accurately determining the pose of a camera mounted on the vehicle relative to an overhead image, such as satellite images, for precise navigation in geographic areas.
Innovation Solution
A computer system uses a machine-learning architecture to generate observed and ground-view feature maps, projecting them geometrically to determine the camera's rotation and translation relative to an overhead image, utilizing multi-head attention layers and neural optimizers for precise pose estimation without requiring multiple images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple images are used for pose determination, then measurement precision is improved, but productivity deteriorates due to increased processing time and computational complexity
Solution Approach 1:
The patent extracts and utilizes geometric relationships between overhead and ground-view images to directly compute pose parameters. By extracting key geometric features and relationships from the images, the system determines camera pose without requiring multiple images, thus maintaining high precision while improving processing efficiency
Solution Approach 2:
The patent creates a projected overhead feature map by projecting the ground-view feature map using geometric relationships. This projected map serves as a computational copy that can be directly compared with the observed overhead feature map to determine pose, eliminating the need for multiple physical images while maintaining accuracy
2Measurement precision
If complex machine-learning architectures are employed, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent introduces a projected overhead feature map as an intermediary between the ground-view image and the overhead image. This intermediary is generated through geometric projection and serves as a bridge for comparison, enabling accurate pose determination through a structured multi-step process that manages system complexity
Solution Approach 2:
The patent segments the pose determination process into distinct stages: generating ground-view and overhead feature maps, projecting the ground-view map to overhead view, comparing the projected and observed overhead maps, and determining pose parameters. This segmentation makes the complex process more manageable and implementable
Data Source
AI summary
A computer includes a processor and a memory, and the memory stores instructions executable by the processor to generate an observed overhead feature map from an overhead image of a geographic area; generate a ground-view feature map from a ground-view image captured by a camera within the geographic area, the camera oriented at least partially horizontally while capturing the ground-view image; project the ground-view feature map to an overhead view based on a geometric relationship, resulting in a projected overhead feature map; determine a rotation of the camera relative to the overhead image based on the projected overhead feature map and the observed overhead feature map; and determine a translation of the camera relative to the overhead image based on the projected overhead feature map and the observed overhead feature map.