Date of Award

3-22-2019

Document Type

Thesis

Degree Name

Master of Science in Electrical Engineering

Department

Department of Electrical and Computer Engineering

First Advisor

Robert C. Leishman, PhD

Abstract

In unmanned aerial navigation the ability to determine the aircraft's location is essential for safe flight. The Global Positioning System (GPS) is the default modern application used for geospatial location determination. GPS is extremely robust, very accurate, and has essentially solved aerial localization. Unfortunately, the signals from all Global Navigation Satellite Systems (GNSS) to include GPS can be jammed or spoofed. To this response it is essential to develop alternative systems that could be used to supplement navigation systems, in the event of a lost GNSS signal. Public and governmental satellites have provided large amounts of high-resolution satellite imagery. These could be exploited through machine learning to aid onboard navigation equipment to provide a geospatial location solution. Deep learning and Convolutional Neural Networks (CNNs) have provided significant advances in specific image processing algorithms. This thesis will discuss the performance of CNN architectures with various hyperparameters and industry leading model designs to address visual aerial localization. The localization algorithm is trained and tested through satellite imagery of a localized area of 150 square kilometers. Three hyper-parameters of focus are: initializations, optimizers, and finishing layers. The five model architectures are: MobileNet V2, Inception V3, ResNet 50, Xception, and DenseNet 201. The hyper-parameter analysis demonstrates that specific initializations, optimizations and finishing layers can have significant effects on the training of a CNN architecture for this specific task. The lessons learned from the hyper-parameter analysis were implemented into the CNN comparison study. After all the models were trained for 150 epochs they were evaluated on the test set. The Xception model with pretrained initialization outperformed all other models with a Root Mean Squared (RMS) error of only 85 meters.

AFIT Designator

AFIT-ENG-MS-19-M-010

DTIC Accession Number

AD1074621

Share

COinS