We've Moved!
Visit SDSU’s new digital collections website at https://digitalcollections.sdsu.edu
Description
There are many situations where temporal aerial imagery can be used to deduce changes in a scene. Two such examples are assessing post-natural-disaster damage and detecting illegal activity along the US-Mexico border region. Automation of aerial image change detection requires computer software that provides automated registration of aerial imagery of a scene. This thesis summarizes work performed to automate aerial image registration. The images and project goals have been provided by the Geography Department of San Diego State University (SDSU) in January 2011 through summer 2012. The images provided have been acquired using a frame center (FC) matching technique originated by the Geography Department of SDSU. The FC matching techniques allows images to be acquired in such a way as to allow for more accurate automated image registration. The automated registration software was developed using the Mathworks product MATLAB (Matrix Laboratory). MATLAB is a simulation environment that allows native manipulation of multi-dimensional data. It provides an interpreted software environment that also includes libraries for image manipulation. This makes it a very good tool for developing and testing an automated image registration algorithm. This is an original algorithm in that it has been designed by using basic image processing principles and routines. The automated image registration algorithm uses concepts such as 2-dimensional normalized cross correlation, feature point matching, and a Random Sample Consensus (RANSAC) like algorithm. The algorithm makes some simplifying assumptions, including using a single, global projective transform for the image registration. The use of a single, global projective transform does not correctly account for variations in terrain. The algorithm provides moderate results, with an average Control Point error of 5 pixels, an average pixel difference of .121 and an average registration time of 8.9 seconds per pair. (These results are averaged over 52 image pairs at 5 sites provided by the Geography Department at SDSU.) The algorithm is provided in text (interpreted) form, but can be compiled into a stand-alone executable. Finally, the algorithm provides batched, near-real-time performance and many diagnostic and understanding tools.
