Aerial Vehicle Tracking With Dynamic Aleatoric Covariance Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
State-of-the-art object detectors lack the ability to provide reliable quantitative measures of uncertainty, particularly aleatoric uncertainty, which is crucial for real-time object tracking in autonomous systems, and existing methods for integrating uncertainty estimation with deep neural networks are computationally expensive or inaccurate.
Innovation Solution
A method and system that utilize a multivariate Gaussian loss function to train a machine learning subsystem to provide full aleatoric covariance matrices for object coordinates and bounding box dimensions in a single pass, followed by a recursive Bayesian estimator to filter noise in real-time using Kalman filters.
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 power consumption increase excessively
Solution Approach 1:
The patent extracts uncertainty estimation from the main DNN inference path by using a separate, lightweight uncertainty module that processes the same features without requiring multiple passes through the full network. This separation allows uncertainty estimation to occur with minimal additional computational overhead, resolving the contradiction between reliability improvement and power consumption increase.
Solution Approach 2:
The patent changes the parameter representation by modeling uncertainty as a covariance matrix derived from a small number of auxiliary outputs rather than through extensive sampling. This parameter transformation enables uncertainty estimation to be computed in a single forward pass, dramatically reducing power consumption while maintaining reliability.
2Reliability
If sampling methods are used to provide uncertainty estimates, then reliability is improved, but processing time increases excessively
Solution Approach 1:
The patent performs preliminary uncertainty estimation during the main inference process by computing covariance matrices from auxiliary outputs that are generated anyway during feature processing. This preliminary action eliminates the need for separate sampling steps, maintaining reliability while avoiding additional processing time delays.
3Productivity
If univariate uncertainty estimation methods are used, then computational efficiency is improved, but measurement precision deteriorates due to inability to capture full covariance
Solution Approach 1:
The patent transitions from univariate to multivariate uncertainty estimation by computing full covariance matrices that capture relationships between multiple output dimensions. This dimensional expansion is achieved efficiently by using auxiliary network outputs that encode correlation information, thereby improving measurement precision without sacrificing computational efficiency.
4Device complexity
If static uncertainty estimation is used with Kalman filters, then device complexity is reduced, but measurement precision deteriorates in the presence of variable noise levels
Solution Approach 1:
The patent introduces dynamic uncertainty estimation by computing covariance matrices that vary with each input frame based on learned features and auxiliary outputs. This dynamic approach adapts to variable noise levels in real-time, improving measurement precision while maintaining relatively low device complexity through efficient matrix computations.
Data Source
Figure 1
Figure 2A
Figure 2B
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.