Landmark-Aided Localization for Air Vehicles Using Learned Object Detectors

Open Access
Author:
DeAngelo, Mark Patrick
Graduate Program:
Aerospace Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
August 09, 2016
Committee Members:
  • Joseph F. Horn, Dissertation Advisor/Co-Advisor
  • Joseph F. Horn, Committee Chair/Co-Chair
  • Robert G. Melton, Committee Member
  • Sean N. Brennan, Committee Member
  • Richard Lee Culver, Outside Member
Keywords:
  • localization
  • feature-based navigation
  • UAV
  • UAS
  • navigation
  • computer vision
  • landmark detection
  • object detection
  • cascade object detector
  • bag of words
  • bag of visual words
  • bag of keypoints
  • SVM
  • visual flight rules
  • VFR
  • image processing
  • UUV
  • AUV
  • sonar
  • underwater
  • airplane
  • quadcopter
  • aircraft
  • unmanned aerial vehicle
  • unmanned aerial system
  • remotely piloted vehicle
  • RPV
  • unmanned underwater vehicle
  • autonomous underwater vehicle
  • localisation
Abstract:
This research presents two methods to localize an aircraft without GPS using fixed landmarks observed from an optical sensor. Onboard absolute localization is useful for vehicle navigation free from an external network. The objective is to achieve practical navigation performance using available autopilot hardware and a downward pointing camera. The first method uses computer vision cascade object detectors, which are trained to detect predetermined, distinct landmarks prior to a flight. The first method also concurrently explores aircraft localization using roads between landmark updates. During a flight, the aircraft navigates with attitude, heading, airspeed, and altitude measurements and obtains measurement updates when landmarks are detected. The sensor measurements and landmark coordinates extracted from the aircraft’s camera images are combined into an unscented Kalman filter to obtain an estimate of the aircraft’s position and wind velocities. The second method uses computer vision object detectors to detect abundant generic landmarks referred as buildings, fields, trees, and road intersections from aerial perspectives. Various landmark attributes and spatial relationships to other landmarks are used to help associate observed landmarks with reference landmarks. The computer vision algorithms automatically extract reference landmarks from maps, which are processed offline before a flight. During a flight, the aircraft navigates with attitude, heading, airspeed, and altitude measurements and obtains measurement corrections by processing aerial photos with similar generic landmark detection techniques. The method also combines sensor measurements and landmark coordinates into an unscented Kalman filter to obtain an estimate of the aircraft’s position and wind velocities.