Convolutional neural network architecture study for aerial visual localization

Abstract

This research studies best practices in convolutional neural network (CNN) architecture design for a novel solution to aerial visual localization. A dataset was developed of satellite images modeling aircraft photographs with corresponding locations in an area of interest. CNN performance, based on varying hyper-parameters such as: optimizers, finishing layers, and weight initializers, was analyzed. Industry leading CNN architectures were trained and tested on the dataset and compared. Finally, a loss function was designed to facilitate returning multiple location estimations for processing with an inertial measurement unit to develop a comprehensive aviation solution.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Mar 21, 2019
Accession Number
AD1074621

Entities

People

  • Jedediah M. Berhold

Organizations

  • Air Force Institute of Technology

Tags

Communities of Interest

  • Autonomy
  • Energy and Power Technologies
  • Space

DTIC Thesaurus Topics

  • Air Force
  • Aircrafts
  • Artificial Intelligence
  • Artificial Intelligence Software
  • Artificial Neural Networks
  • Cameras
  • Computer Languages
  • Computer Programming
  • Computer Vision
  • Computers
  • Convolutional Neural Networks
  • Coordinate Systems
  • Global Navigation Satellite Systems
  • Global Positioning Systems
  • Governments
  • Image Processing
  • Inertial Measurement Units
  • Information Science
  • Kalman Filters
  • Machine Learning
  • Navigation
  • Neural Networks
  • Operating Systems
  • Unmanned Aerial Vehicles
  • World Geodetic System

Readers

  • Neural Network Machine Learning.
  • Sensor Fusion and Tracking Systems.

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks
  • Space