Stereo Visual Odometry Integrity Monitoring for Fault Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vision-aided navigation systems for safety-critical applications like Urban Air Mobility and autonomous vehicles face challenges in ensuring the integrity, reliability, and accuracy of visual odometry measurements, particularly in environments where Global Navigation Satellite System (GNSS) is unavailable.
Innovation Solution
A method for integrity monitoring of visual odometry using stereo vision sensors, which includes image preprocessing, feature extraction, temporal feature matching with a mismatching limiting discriminator, depth recovery with a range error limiting discriminator, outlier rejection with a modified RANSAC technique, and state vector estimation with a solution separation technique to determine changes in rotation and translation, detect faults, and compute protection levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing visual odometry approaches are used to improve accuracy and robustness of measurements, then measurement quality is improved, but integrity and reliability of navigation measurements deteriorate due to lack of rigorous error characterization
Solution Approach 1:
The patent transforms the visual odometry system by introducing integrity monitoring parameters including error magnitude characterization, fault probability assessment, and protection level computation. These new parameters enable rigorous reliability evaluation while maintaining measurement accuracy through the solution separation technique and overbounding Gaussian models.
2Reliability
If multiple discriminator techniques are applied to reduce feature mismatch rates and limit range errors, then measurement reliability is improved, but system complexity increases due to additional processing steps
Solution Approach 1:
The patent segments the integrity monitoring process into distinct functional modules: temporal feature matching with mismatching limiting discriminator, range recovery with range error limiting discriminator, outlier rejection with modified RANSAC, and state vector estimation with solution separation. This segmentation organizes complexity into manageable components while achieving comprehensive reliability monitoring.
Solution Approach 2:
The patent applies preliminary discrimination actions before final state estimation by using mismatching limiting discriminators to pre-filter feature matches, range error limiting discriminators to pre-validate depth measurements, and modified RANSAC to pre-reject outliers. This preliminary filtering reduces the burden on subsequent integrity monitoring steps.
3Reliability
If solution separation technique is used to detect faults and compute protection levels, then integrity monitoring capability is improved, but computational requirements increase
Solution Approach 1:
The patent applies local quality optimization by using overbounding Gaussian models to characterize error magnitudes and fault probabilities in localized measurement spaces. This approach computes protection levels for specific state variables (rotation and translation changes) rather than globally, reducing computational requirements while maintaining rigorous integrity monitoring capability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method of integrity monitoring for visual odometry comprises capturing a first image at a first time epoch with stereo vision sensors, capturing a second image at a second time epoch, and extracting features from the images. A temporal feature matching process is performed to match the extracted features, using a feature mismatching limiting discriminator. A range, or depth, recovery process is performed to provide stereo feature matching between two images taken by the stereo vision sensors at the same time epoch, using a range error limiting discriminator. An outlier rejection process is performed using a modified RANSAC technique to limit feature moving events. Feature error magnitude and fault probabilities are characterized using overbounding Gaussian models. A state vector estimation process with integrity check is performed using solution separation to determine changes in rotation and translation between images, determine error statistics, detect faults, and compute protection level or integrity risk.