Visual Feature Database Construction via 3D Model Projection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional visual positioning methods are limited by the small amount of information gathered from feature points in the visual feature database due to the fixed number of database creation images, leading to inaccurate positioning, especially in scenarios with large radiation differences and weak textures.
Innovation Solution
The method involves intersecting a ray corresponding to a feature point from a database creation image with a 3D model to determine its 3D position, allowing for the inclusion of more feature points in the visual feature database, which includes semantic information and confidence degrees, and using scene simulation and segmentation to enhance feature extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If feature matching is performed on captured images to construct the visual feature database, then the database includes information about successfully matched feature points, but the quantity of feature points in the database is limited due to the fixed number of database creation images
Solution Approach 1:
Instead of extracting 3D positions from multiple 2D images through feature matching, the patent inverts the approach by projecting 3D model points onto 2D images to generate synthetic feature points. This allows the system to overcome the limitation of fixed database creation images by generating unlimited feature points through 3D model projections.
Solution Approach 2:
The patent creates synthetic feature points by copying and projecting 3D model points onto 2D images. These synthetic feature points are generated based on the 3D model and camera parameters, allowing the database to contain far more feature points than would be available from actual image capture alone.
2Productivity
If the quantity of database creation images is fixed, then the database construction process is efficient, but the amount of information about feature points in the visual feature database is comparatively small
Solution Approach 1:
The patent performs preliminary action by pre-establishing a 3D model of the scene before constructing the visual feature database. This 3D model serves as a foundation for generating synthetic feature points, allowing the system to enrich the database without capturing additional images, thus preventing information loss while maintaining construction efficiency.
Solution Approach 2:
The system copies 3D model points and projects them onto 2D images to create synthetic feature points. This copying process generates additional feature point information without requiring more database creation images, thereby preserving information completeness while maintaining productivity.
3Measurement precision
If conventional feature matching is used to obtain 3D positions of feature points, then only successfully matched feature points can be saved, but this limits the quantity of feature points that can be gathered from a fixed number of images
Solution Approach 1:
The patent inverts the conventional approach by not trying to infer 3D positions from 2D images through matching, but instead projecting known 3D model points onto 2D images to create synthetic feature points. This inversion allows the system to achieve both accurate 3D positions (from the known 3D model) and a large quantity of feature points (through systematic projection).
Solution Approach 2:
The system copies 3D model points with known precise positions and projects them onto 2D images. This ensures that each synthetic feature point inherits the accurate 3D position information from the source model, maintaining measurement precision while dramatically increasing the quantity of available feature points.
Data Source
AI summary
This application provides a visual feature database construction method, including: obtaining a database creation image; performing feature extraction on the database creation image to obtain a feature point of the database creation image and a descriptor of the feature point of the database creation image; intersecting a ray corresponding to the feature point of the database creation image with a 3D model, to determine the 3D position of the intersection point at which the ray intersects with the 3D model as the 3D position of the feature point of the database creation image; and writing the descriptor of the feature point of the database creation image and the database creation image into a visual feature database, to construct the visual feature database.