Date of Award

1-1-2022

Publication Type

Thesis

Degree Name

M.A.Sc.

Department

Electrical and Computer Engineering

Keywords

Global Position System, Unmanned Aerial Vehicles, Computationally intensive computer vision

Rights

info:eu-repo/semantics/embargoedAccess

Abstract

Unmanned Aerial Vehicles (UAVs) depend on Global Position System (GPS) for determining their own location during navigation. In GPS-denied environments, a UAV needs to make use of alternative strategies for location estimation. Computer vision and machine learning algorithms can be used to detect common landmarks such as buildings, trees, and road intersections from aerial views. Landmark detection in combination with geotagging can be utilized for UAV self-localization. Graphical Processing Units (GPUs) such as Nvidia's Jetson have shown great promise for accelerating computationally intensive computer vision and machine learning algorithms. This thesis presents a novel method for an optimized GPU implementation of a deep neural network algorithm for landmark detection to estimate UAV's location in GPS-denied environments. It compares Nvidia's Jetson Nano, TX2 and AGX Xavier in terms of their performance in landmark detection and recognition. Darknet and ResNet frameworks have been implemented for image processing using Convolutional Neural Networks(CNN). Processing time and numerical accuracy are measured on existing image datasets such as DOTA, UC Merced Land Use and PatternNet, and on new image datasets obtained from the cameras placed on-board a real UAV. Results demonstrate that the proposed vision-based self-localization of UAV offers a promising solution to navigation in a GPS-denied Environment.

Share

COinS