Object Pose Recognition Using 3D Point Cloud Iterative Shifting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object pose recognition systems lack robustness against environmental variations such as illumination changes and geometric transformations, making them impractical for robot services that require recognizing objects in diverse settings.
Innovation Solution
An object pose recognition apparatus and method that shifts previously stored 3D point cloud data based on calculated feature points and recalculates the object pose using input image data, repeating the process until convergence conditions are met, and restores unrestored pixels using adjacent brightness values from 3D point cloud data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional object pose recognition methods are used, then the system is simple to implement, but the robustness against environmental variations (illumination changes, geometric transformations) deteriorates
Solution Approach 1:
The patent transforms 2D image data into 3D point cloud data by introducing a depth dimension. This dimensional enhancement allows the system to capture geometric information that is invariant to illumination changes and provides more robust feature matching across different viewing angles and distances, directly improving robustness against environmental variations
Solution Approach 2:
The patent performs preliminary registration and alignment of 3D point cloud data before final pose recognition. By pre-processing the 3D data to establish a reference coordinate system and pre-align multiple views, the system creates a more stable foundation for pose estimation that is less sensitive to environmental variations during actual recognition
2Reliability
If 3D point cloud data is shifted and pose is recalculated iteratively, then the invariance performance improves, but the processing time increases
Solution Approach 1:
The patent implements an iterative refinement process where the pose is calculated, the 3D point cloud is shifted based on the calculated pose, and the pose is recalculated using the shifted data. This periodic iteration continues until convergence conditions are met, progressively improving invariance performance by repeatedly refining the pose estimation
Solution Approach 2:
The patent uses feedback from the pose calculation results to guide the shifting of 3D point cloud data. The calculated pose serves as feedback that determines how the 3D data should be transformed, and this transformed data is then used to recalculate the pose, creating a closed-loop system that converges to an accurate and invariant pose estimation
3Measurement precision
If feature points are extracted and matched from multiple image data, then the pose calculation accuracy improves, but the computational complexity increases
Solution Approach 1:
The patent extracts key feature points from both 2D image data and 3D point cloud data, isolating the most informative elements for pose calculation. By selecting only salient feature points rather than processing all pixels or data points, the system achieves accurate pose estimation while reducing computational complexity
Solution Approach 2:
The patent uses feature points that serve multiple functions: they are used for initial pose estimation, for matching between 2D and 3D data, and for verifying the accuracy of pose calculation. This multi-functional use of feature points maximizes the information extracted from the data while avoiding redundant computational steps
Data Source
AI summary
An object pose recognition apparatus and method. The object pose recognition method includes acquiring first image data of an object to be recognized and 3-dimensional (3D) point cloud data of the first image data, and storing the first image data and the 3D point cloud data in a database, receiving input image data of the object photographed by a camera, extracting feature points from the stored first image data and the input image data, matching the stored 3D point cloud data and the input image data based on the extracted feature points and calculating a pose of the photographed object, and shifting the 3D point cloud data based on the calculated pose of the object, restoring second image data based on the shifted 3D point cloud data, and re-calculating the pose of the object using the restored second image data and the input image data.