3D Image Patch Extraction for Robot Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Simultaneous localization and map-building (SLAM) technologies face challenges in accurately localizing robots and building maps due to errors caused by changes in the spatial relationship between the camera and feature points in 3D images, as existing methods extract image patches of a fixed size, leading to inaccuracies in position recognition.
Innovation Solution
A method involving the extraction of feature points from 3D images, conversion of planes including these points into reference images, and matching image patches using an iterative closest point (ICP) algorithm, with the inclusion of brightness value analysis to ensure accurate 3D surface information matching and minimize errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image patches of fixed size are extracted from 3D images for SLAM, then the extraction process is simple, but localization accuracy deteriorates due to errors caused by changes in spatial relationship between camera and feature points
Solution Approach 1:
The patent changes the parameter of image patch extraction by transitioning from fixed-size patches to patches with adaptive size and shape. Specifically, the system extracts patches based on the actual 3D geometry of feature points, allowing patch dimensions to vary according to the spatial relationship between the camera and feature points. This parameter adaptation resolves the contradiction by maintaining localization accuracy while managing processing complexity through intelligent rather than brute-force methods.
Solution Approach 2:
The patent introduces a new dimension to the image patch extraction process by incorporating 3D spatial information into the 2D image processing. The system uses depth information and 3D coordinates to guide the extraction of image patches, transforming the problem from a purely 2D image processing task to a 3D-aware process. This dimensional enhancement allows accurate localization despite changes in camera-to-feature-point spatial relationships.
2Measurement precision
If image patches are extracted without considering 3D position information, then the processing is faster, but map-building accuracy deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing 3D position information of feature points before the actual image patch extraction and matching process. The system performs preliminary 3D reconstruction and feature point localization, then uses this pre-computed information to guide subsequent image processing steps. This preliminary preparation ensures map-building accuracy while reducing the computational burden during real-time operation, thus addressing the time-accuracy tradeoff.
Solution Approach 2:
The patent substitutes mechanical brute-force processing with an intelligent system that leverages 3D geometric information. Instead of exhaustively processing all image regions or using computationally intensive methods, the system uses 3D position data to intelligently select and process only relevant image patches. This substitution of mechanical processing with geometry-guided processing maintains accuracy while significantly reducing processing time.
Data Source
AI summary
Disclosed herein are a feature point used to localize an image-based robot and build a map of the robot and a method of extracting and matching an image patch of a three-dimensional (3D) image, which is used as the feature point. It is possible to extract the image patch converted into the reference image using the position information of the robot and the 3D position information of the feature point. Also, it is possible to obtain the 3D surface information with the brightness values of the image patches to obtain the match value with the minimum error by a 3D surface matching method of matching the 3D surface information of the image patches converted into the reference image through the ICP algorithm.