Automatic UAV image registration using feature detection and matching with satellite imagery

Abstract

This thesis presents an approach and implementation for using satellite imagery to perform image registration with unprocessed aerial images captured with a UAV. Aerial imagery is used in a large variety of applications including disaster relief, urban planning, crop and vegetation monitoring, and mapping. The common difficulty in utilizing aerial imagery captured by aircraft rather than satellite is that of image registration: transforming data into the image coordinate system. Often, the goal of these applications involves transforming location data into the image coordinate system for object extraction and further processing. Current approaches require hand-labeling of correspondences, the use of ground control points (GCPs), or human analysis to identify objects or locations of interest in the aerial imagery. As such, these methods do not sufficiently scale, generalize, or provide the efficiency required for these applications. The proposed approach to image registration in aerial imagery makes use of raw images captured on consumer-grade UAVs, and uses automatic feature detection and matching to register the imagery. This implementation provides a proof of concept, which was found to succeed on roughly 65% of our images.