Stereo Camera Position Estimation Using Learning Model Abnormality Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current techniques for estimating the position/orientation of image capturing devices, such as those used in autonomous driving, lack stability and accuracy, particularly in noisy environments or with insufficient training data, leading to potential errors in self-position estimation.
Innovation Solution
An information processing device employing a stereo configuration with two image capturing devices, utilizing a learning model to estimate geometric information and position/orientation, which combines provisional information from multiple sources to generate stable position/orientation data, and includes mechanisms to detect and mitigate issues such as noise and abnormality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single image capturing device is used for position/orientation estimation, then the device complexity is reduced, but the measurement precision and reliability deteriorate
Solution Approach 1:
The patent combines multiple image capturing devices (first and second image capturing devices with overlapping visual fields) to capture multiple images of the same scene. By merging the information from these multiple devices and images, the system achieves more reliable and precise position/orientation estimation than would be possible with a single device, while managing the complexity through systematic processing of the combined data
2Productivity
If a learning model is used to estimate geometric information, then the productivity is improved, but the reliability deteriorates in noisy environments or with insufficient training data
Solution Approach 1:
The patent implements a feedback mechanism where the system detects abnormalities in the geometric information estimated by the learning model. When noise or insufficient training data causes unreliable estimates, the abnormality detection unit identifies these issues and the system can request re-input of images or adjust processing, thereby maintaining reliability while preserving the productivity benefits of using a learning model
Solution Approach 2:
The patent performs preliminary actions by detecting abnormalities in the estimated geometric information before final position/orientation calculation. This preliminary detection allows the system to identify and mitigate issues caused by noise or insufficient training data, ensuring that only reliable geometric information is used for the final estimation, thus maintaining both productivity and reliability
3Measurement precision
If multiple provisional geometric information estimates are combined, then the measurement precision is improved, but the device complexity increases
Solution Approach 1:
The patent segments the position/orientation estimation process into distinct functional units: multiple image capturing devices capture images independently, the learning model estimates geometric information from each image separately, the abnormality detection unit checks each estimate independently, and finally the generating unit combines the validated estimates. This segmentation allows the system to achieve high measurement precision through multiple estimates while managing complexity through modular, organized processing
Data Source
AI summary
An information processing device finds a position/orientation of an image capturing unit including first and second cameras whose image capturing fields are partially overlapped, the information processing device comprises an input unit that inputs first and second images from the first and second cameras a holding unit holding a learning model; a first estimating unit that estimates first provisional geometric information from the first and second images; a second estimating unit that estimates second provisional geometric information on the basis of the first image and the learning model; a third estimating unit that estimates third provisional geometric information on the basis of the second image and the learning model; and a generating unit that, on the basis of the first, second, and third provisional geometric information, generates the position/orientation information of the image capturing unit.