3D Object Attitude Estimation Using Illumination-Invariant Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating the attitude of objects in a three-dimensional space face challenges such as reduced accuracy due to changes in illumination conditions, the need for capturing points on the object's surface, and the difficulty of using regression models with defective input data, especially when multiple objects cannot be discriminated from each other.
Innovation Solution
The proposed solution involves a multi-step process using a computer to calculate representative points from images obtained by multiple imaging devices, estimate the positions of objects in 3D space, extract feature amounts from image regions, and use a preliminarily learned regression model to estimate the attitudes of objects based on their positions and feature amounts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If images themselves are used for attitude estimation, then the method is simple to implement, but the estimation accuracy is reduced by changes in illumination condition
Solution Approach 1:
The patent extracts specific feature quantities from images that are invariant to illumination changes, such as contour shapes and geometric properties, rather than using raw image data. This extraction process isolates the essential attitude-related information while removing illumination-dependent variations.
Solution Approach 2:
The patent transforms image data into different parameter representations (feature quantities) that are less sensitive to illumination changes. By changing from raw pixel values to extracted geometric and structural features, the system achieves illumination invariance while maintaining attitude estimation capability.
2Measurement precision
If points on the surface of a three-dimensional object are captured, then the attitude can be calculated using point positions, but the attitude cannot be estimated when these points cannot be observed
Solution Approach 1:
The patent creates a universal attitude estimation system that works with multiple types of input data (images, point clouds, feature quantities) and can handle various observation conditions. The system can estimate attitude whether surface points are directly observable or only indirect features are available, making the method universally applicable.
Solution Approach 2:
The patent introduces feature quantities as an intermediary between raw observations and attitude estimation. These features serve as mediators that can be derived from various sources (images, partial point clouds) and reliably indicate attitude even when direct surface point observation is impossible.
3Adaptability or versatility
If a regression model is used for attitude estimation, then the method can be learned from pair data, but it is difficult to achieve when part of input data includes a defect
Solution Approach 1:
The patent performs preliminary processing of input data to identify and handle defective regions before attitude estimation. By pre-processing images and point clouds to detect defects and compensate for missing information, the system ensures reliable input data for the regression model even when original observations are imperfect.
Solution Approach 2:
The patent prepares multiple complementary data sources and feature extraction methods in advance to cushion against potential data defects. If one input source is defective, alternative sources or feature representations can compensate, ensuring the regression model always receives sufficient valid information.
4Reliability
If multiple regression models are prepared for different defect situations, then the attitude can be estimated even with defective input data, but the system complexity increases
Solution Approach 1:
The patent develops a single universal regression model that can handle various defect situations through flexible feature engineering and data preprocessing. Rather than creating multiple specialized models, one robust model processes all cases by adapting to the quality and type of available input data.
Solution Approach 2:
The patent changes the parameter representation of input data dynamically based on defect detection, transforming defective inputs into suitable feature formats that the single regression model can process effectively. This parameter adaptation allows one model to replace multiple specialized models.
Data Source
AI summary
An estimation step according to an embodiment causes a computer to execute: a calculation step of using a plurality of images obtained by a plurality of imaging devices imaging a three-dimensional space in which a plurality of objects reside, to calculate representative points of pixel regions representing the objects among pixel regions of the images; a position estimation step of estimating positions of the objects in the three-dimensional space, based on the representative points calculated by the calculation step; an extraction step of extracting predetermined feature amounts from image regions representing the objects; and an attitude estimation step of estimating attitudes of the objects in the three-dimensional space, through a preliminarily learned regression model, using the positions estimated by the position estimation step, and the feature amounts extracted by the extraction step.