Stereo-IMU Navigation State Augmentation With Minimal Error States
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robotic navigation systems face challenges in providing real-time state estimation with high accuracy and robustness, especially in environments where GPS is unavailable or inaccurate, due to the computational complexity of tightly-coupled stereo-IMU SLAM algorithms, which are not suitable for resource-limited processors and drift significantly.
Innovation Solution
The MAVeN-stereo algorithm uses a tightly-coupled Extended Kalman Filter estimator that requires only six extra error states to process multiple stereo cameras and an IMU, allowing for real-time state estimation without a pre-mapped terrain model, reducing computational cost and drift, and maintaining accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If tightly-coupled stereo-IMU SLAM algorithms are used to improve measurement precision and robustness, then navigation accuracy is improved, but computational complexity and runtime increase significantly
Solution Approach 1:
The patent segments the state vector into two parts: a fixed 15-state inertial navigation component and a variable number of visual features. By treating visual features as measurements rather than state variables, the core estimator maintains constant size (15 states) independent of the number of features, thus reducing computational complexity while maintaining tight coupling benefits
Solution Approach 2:
The patent extracts visual features from the state vector and treats them as external measurements. This removes the computational burden of estimating feature positions as state variables, allowing the system to maintain high measurement precision through tight coupling while reducing the device complexity by eliminating feature state estimation
2Measurement precision
If tightly-coupled stereo-IMU SLAM algorithms are used to improve measurement precision, then navigation accuracy is improved, but software runtime increases making it unsuitable for resource-limited processors
Solution Approach 1:
The patent segments the estimation problem into a fixed inertial navigation core (15 states) and variable visual measurements. This segmentation allows the computationally intensive tight-coupled estimation to operate on a fixed-size state vector, significantly reducing software runtime while maintaining navigation accuracy through the preserved tight coupling between IMU and stereo vision data
3Productivity
If prior art autonomous lunar rover VO is used to reduce computational cost, then runtime is reduced, but position drift increases to the order of 54%
Solution Approach 1:
The patent implements feedback by using the fixed 15-state inertial navigation model to continuously correct drift through tight coupling with stereo vision measurements. The estimator uses reprojection error feedback to maintain position accuracy while operating with reduced computational cost compared to full SLAM approaches
Solution Approach 2:
The patent changes the parameter representation by using a fixed 15-state inertial navigation state vector instead of variable-size SLAM state vectors. This parameter change enables reduced computational cost while maintaining reliability through the preserved inertial navigation model that inherently handles drift correction
4Measurement precision
If each visual feature is included in the state vector of the estimator, then measurement precision is improved, but the number of error states increases leading to large numerical complexity
Solution Approach 1:
The patent segments the estimation problem by separating inertial navigation states (fixed 15 states) from visual features (variable measurements). This segmentation maintains measurement precision by keeping visual features in the estimation process while reducing device complexity by eliminating feature states from the error vector, resulting in a constant 15-state error vector independent of feature count
Data Source
AI summary
A method to perform vision based navigation on a moving vehicle. An image processor receives reference images with associated depth information as base frames or search frames. 2D features are detected in base frames and tracked in subsequent search frames. A new base frame is triggered when a number of features or a spatial distribution fall under thresholds. A depth processor reconstructs the depth and 3D position of each of the 2D features. A track manager manages a feature list database mapping the 2D coordinates of each tracked 2D feature with the 3D coordinate of that feature. A state manager constructs a filter state vector with 15 error states propagated at an IMU rate and 6 additional error states corresponding to clones of pose states. A visual updater utilizes the tracked 2D feature coordinates to update the filter state vector. A filter forms residuals and corrects inertial error drift.