DEEP LEARNING MODEL UNCERTAINTY ESTIMATION USING MONTE CARLO DROPOUT TO IMPROVE UAV TARGET POSE PREDICTIONS
Loading...
Authors
Hooker, Alex E.
Subjects
artificial intelligence
machine learning
deep learning
pose estimation
directed energy weapon
out-of-distribution data
classification-regression
uncertainty estimation
epistemic uncertainty
Monte Carlo dropout
machine learning
deep learning
pose estimation
directed energy weapon
out-of-distribution data
classification-regression
uncertainty estimation
epistemic uncertainty
Monte Carlo dropout
Advisors
Agrawal, Brij N.
Kim, Jae Jun
Date of Issue
2024-06
Date
Publisher
Monterey, CA; Naval Postgraduate School
Language
Abstract
Target pose estimation is an important function of many relevant military applications including satellite rendezvous operations and directed energy weapons systems. This task is best left to deep learning (DL) models that can react more quickly than human operators; however, epistemic uncertainty exists in every model, which is manifested in the predictions it makes on inputs outside of its training distribution. This thesis develops a methodology to inform users whether a model’s output pose prediction should be trusted based on the model’s uncertainty. Model uncertainty was quantified using Monte Carlo dropout on in-distribution and out-of-distribution images of unmanned aerial vehicles (UAVs). Images of miniature UAVs taken using a laser beam testbed were passed through the model multiple times, each time producing a set of pose predictions due to the stochastic nature of dropout. The variances of the prediction sets were computed and were the proxy for model uncertainty. At inference, model error was measured as the geodesic distance between each ground truth pose label and the mean prediction in each set. Through experimentation, a correlation between model error and model uncertainty was identified and used to propose a threshold of variance above which the model was not to be trusted.
Type
Thesis
Description
Series/Report No
Organization
Identifiers
NPS Report Number
Sponsors
N/A
Funder
Format
Citation
Distribution Statement
Distribution Statement A. Approved for public release: Distribution is unlimited.
Rights
This publication is a work of the U.S. Government as defined in Title 17, United States Code, Section 101. Copyright protection is not available for this work in the United States.