Aerial Vehicle Tracking With Dynamic Aleatoric Uncertainty Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing single-pass object detectors, such as those used in autonomous systems, lack the ability to provide reliable quantitative measures of uncertainty, particularly aleatoric uncertainty, which is crucial for real-time tracking of high-speed objects and integration with larger decision-making systems, and existing methods for uncertainty estimation are computationally expensive or inaccurate.
Innovation Solution
A method and system that utilize a multivariate deep evidential regression network trained with a Gaussian loss to provide full aleatoric uncertainty covariance matrices for object coordinates and bounding box dimensions, combined with a recursive Bayesian estimator using Kalman filters to filter noise in real-time, allowing for accurate and efficient tracking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If sampling methods (Bayesian neural networks, ensemble methods, Monte Carlo dropout) are used to estimate uncertainty in DNNs, then uncertainty estimation capability is improved, but computational cost and processing time increase excessively
Solution Approach 1:
The patent extracts the uncertainty estimation function from the main DNN processing pipeline by using a separate, lightweight uncertainty module that operates independently. This allows the main detection network to run efficiently while uncertainty is estimated through a different, less computationally intensive pathway that doesn't require multiple sampling passes through the entire network.
Solution Approach 2:
The patent employs a simplified, computationally inexpensive uncertainty estimation approach that sacrifices some of the sophistication of full sampling methods but provides sufficient uncertainty information for real-time tracking. This 'cheap' estimation method is designed to be disposable in terms of computational resources, allowing rapid execution without the heavy overhead of traditional sampling-based uncertainty quantification.
2Reliability
If sampling methods are used to provide uncertainty estimates, then reliability of uncertainty measurement is improved, but device complexity and processing footprint increase
Solution Approach 1:
The uncertainty estimation functionality is extracted from the complex sampling-based approaches and implemented through a streamlined mechanism that operates with minimal processing footprint. This extracted uncertainty module provides reliable measurements without requiring the full infrastructure of sampling-based methods.
Solution Approach 2:
The patent replaces the mechanical sampling process (repeated forward passes through the network) with a more efficient computational mechanism that achieves similar uncertainty quantification goals without the same hardware resource demands. This substitution reduces the processing footprint while maintaining measurement reliability.
3Speed
If single-pass object detectors are used for real-time tracking, then processing speed is improved, but aleatoric uncertainty estimation capability deteriorates
Solution Approach 1:
The patent merges the single-pass detection capability with uncertainty estimation by integrating a lightweight uncertainty module that works in conjunction with the fast detector. This combination allows the system to maintain the speed advantages of single-pass processing while adding the previously missing uncertainty estimation capability through a complementary, non-interfering mechanism.
4Ease of manufacture
If univariate normal inverse-gamma distribution is used for uncertainty estimation, then computational simplicity is improved, but ability to capture full aleatoric uncertainty covariance deteriorates
Solution Approach 1:
The patent changes the parameter representation from univariate variance estimates to a covariance-based formulation that captures the full aleatoric uncertainty structure. This parameter change allows the system to maintain computational efficiency while accurately representing the correlated uncertainties in object detection, thereby resolving the limitation of treating bounding box dimensions as independent variables.
Data Source
AI summary
Techniques for aerial vehicle tracking using dynamic aleatoric uncertainty covariance estimation are presented. The techniques include: obtaining an image depicting at least one aerial vehicle of interest; passing the image to a first machine learning subsystem, which provides at least one feature vector; inputting the at least one feature vector to a second machine learning subsystem, where the second machine learning subsystem is trained to provide detected aerial vehicle identification data sets (including respective aerial vehicle coordinates, respective aerial vehicle bounding box dimensions, and respective dynamic aleatoric uncertainty covariance values) corresponding to input feature vectors; providing at least one detected aerial vehicle identification data set to a recursive Bayesian estimator subsystem, from which at least one filtered set of aerial vehicle coordinates, representing a real-time location of a respective aerial vehicle of interest, is obtained; and outputting the at least one filtered set of aerial vehicle coordinates.