Terrain Image Localization for Aerial Vehicles With Dynamic Uncertainty
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing GPS-denied localization methods for aerial vehicles face challenges in accurately determining location due to variable noise levels and lack of reliable uncertainty estimation, particularly in visual place recognition, leading to inaccurate results from static uncertainty models and computationally expensive sampling methods.
Innovation Solution
A system combining an evidential deep learning neural network with a recursive Bayesian estimator uses dynamic uncertainty estimates to improve localization accuracy by providing both similarity and uncertainty values in real-time, leveraging a single pass through the neural network and a particle filter for reweighting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If hand-tuned measurement models with static uncertainty estimates are used, then the system is simple to implement, but localization accuracy deteriorates due to inability to account for variable noise levels
Solution Approach 1:
The patent implements dynamic uncertainty estimation by training the neural network to output uncertainty values that adapt to each input image's noise characteristics. The measurement model transitions from static to dynamic, allowing uncertainty to vary with each measurement based on image quality, lighting conditions, and feature distinctiveness.
Solution Approach 2:
The patent changes the parameter of uncertainty from a fixed static value to a dynamic variable that changes with each input image. The neural network learns to predict uncertainty parameters based on image characteristics, enabling the system to adapt to varying noise levels without increasing implementation complexity.
2Reliability
If sampling methods (Bayesian neural networks, ensemble methods, Monte Carlo dropout) are used for uncertainty estimation, then measurement reliability improves, but computational cost and power consumption increase excessively
Solution Approach 1:
The patent replaces expensive sampling-based uncertainty estimation with a cheaper alternative: a single forward pass through a trained neural network that directly outputs uncertainty values. This disposable approach achieves reliable uncertainty estimation without the computational burden of multiple sampling passes.
Solution Approach 2:
The patent substitutes the mechanical sampling process (repeated passes through the network) with a learned direct estimation mechanism. The neural network is trained to predict uncertainty directly, replacing the iterative sampling mechanism with a single computational pass that consumes significantly less power.
3Reliability
If sampling methods are used for uncertainty estimation, then measurement reliability improves, but processing time and computational footprint increase excessively
Solution Approach 1:
The patent performs preliminary training of the neural network offline to learn the relationship between image characteristics and uncertainty. During runtime, the pre-trained network directly outputs uncertainty values without requiring sampling, having already performed the computationally intensive work during the training phase.
Solution Approach 2:
The patent replaces time-consuming sampling operations with a single forward pass through the trained neural network. This disposable approach provides reliable uncertainty estimation in real-time without the computational overhead of multiple sampling iterations.
4Device complexity
If static uncertainty estimates are used in statistical estimators, then computational complexity is reduced, but measurement precision deteriorates due to flattening of variable noise levels
Solution Approach 1:
The patent changes the uncertainty parameter from static to dynamic by having the neural network predict uncertainty values specific to each input image. This allows the measurement model to account for variable noise levels while maintaining computational efficiency through a single forward pass.
Solution Approach 2:
The patent introduces dynamics into the uncertainty estimation by making it adaptive to each measurement. The neural network learns to adjust uncertainty values based on image characteristics, enabling precise localization without increasing computational complexity significantly.
Data Source
AI summary
Techniques for localizing a vehicle in real time using dynamic uncertainty estimates are presented. The techniques include obtaining a terrain image captured by the vehicle; passing the terrain image to a trained evidential deep learning neural network subsystem, from which a dynamic uncertainty value and a first feature vector are obtained in real time; for each of a plurality of candidate terrain locations, comparing the first feature vector to a respective second feature vector representative of a candidate terrain location, from which a respective similarity score is obtained; for at least one of the plurality of candidate terrain locations, updating in real time, by a recursive Bayesian estimator, a respective location weight based on the dynamic uncertainty value and the respective similarity score; estimating, in real time, a location of the vehicle based on the plurality of location weights; and providing the location of the vehicle.