Robot Re-Localization Using HF-Net Keyframe Feature Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robot localization systems face challenges in re-localization, particularly in changing scenes, with Laser SLAM systems relying on resource-intensive 3D maps and Visual SLAM systems requiring costly 3D map establishment, leading to suboptimal performance.
Innovation Solution
A method using a Hierarchical Feature Network (HF-Net) model for feature extraction from RGBD images, storing image features and poses of keyframes without creating a full 3D map, and employing a 2D-2D pose estimation method for re-localization, allowing efficient re-localization even in changing environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If L-SLAM system uses submaps for re-localization by mapping current submap to stored submaps, then re-localization can be achieved, but resource consumption increases due to storing and processing multiple 2D grid submaps
Solution Approach 1:
The patent extracts only the essential 2D image features and pose information from keyframes, discarding the need to store complete 3D maps or extensive 2D grid submaps. This selective extraction maintains re-localization capability while significantly reducing resource consumption by storing only compressed feature descriptors and transformation matrices.
Solution Approach 2:
The patent creates simplified 2D feature representations (descriptors) that copy only the essential visual characteristics needed for matching, rather than storing complete 3D map data or full submaps. These compressed feature copies enable efficient comparison and re-localization with minimal storage and processing requirements.
2Measurement precision
If V-SLAM system uses Bag of Words method with 3D landmarks and PNP algorithm for re-localization, then pose calculation can be performed, but resource overhead increases due to 3D map establishment and triangulation
Solution Approach 1:
The patent extracts only 2D feature descriptors and corresponding pose information from keyframes, eliminating the need for complex 3D triangulation and landmark establishment. This extraction approach maintains sufficient accuracy for re-localization while dramatically simplifying the system by working directly with 2D image features and transformation matrices.
Solution Approach 2:
The patent replaces the complex mechanical process of 3D triangulation, point cloud generation, and spatial reconstruction with a simpler 2D feature matching approach using descriptor comparison and pre-stored pose transformations. This substitution eliminates the need for computationally intensive 3D geometry calculations while maintaining re-localization functionality.
3Loss of information
If existing systems store complete 3D maps for re-localization, then comprehensive environmental information is available, but storage requirements and processing time increase significantly
Solution Approach 1:
The patent extracts only the critical elements needed for re-localization: 2D feature descriptors, keyframe poses, and transformation matrices. This selective extraction preserves sufficient environmental information for accurate re-localization while eliminating redundant data, thereby reducing both storage requirements and processing time for matching operations.
Data Source
AI summary
A method for re-localization of the robot may include retrieving, for each of keyframes in a keyframe database of the robot, image features and a pose of the keyframe, the image features of the keyframe comprising a global descriptor and local descriptors of the keyframe (210); extracting image features of a current frame captured by the robot, the image features of the current frame comprising a global descriptor and local descriptors of the current frame (220); determining one or more rough matching frames from the keyframes based on comparison between the global descriptor of each keyframe and the global descriptor of the current frame (230); determining a final matching frame from the one or more rough matching frames based on comparison between the local descriptors of each rough matching frame and the local descriptors of the current frame (240); and calculating a pose of the current frame based on a pose of the current frame based on a pose of the final matching frame (250).