Growth in the prevalence of cameras has resulted in larger amounts of available image data. This has resulted in demand for automated methods of analysing this data. One key area of demand is automated change detection, the automated detection of changes in a scene, as recorded by a reference and sample image. Established methods of change detection tend to rely on the reference and sample image being captured from the same position, but much of the available data does not fit this criteria. This thesis presents novel approaches to key challenges in wide-baseline cases involving differences in viewing angle of up to 30 degrees, including registration and the image region matching that are robust to the inherent registration errors. The developed algorithms are then combined into an end-to-end system.

This thesis presents novel registration approaches including the use of a Delaunay triangulation mask that enables registration of each component triangle, a method of finding local planes in scenes by clustering matched feature points, the use of edge detection to register the edges of objects, and a method for registering planes that are orthogonal to a defined image plane and to the camera line. These techniques allow for the registration of complex 3D scenes with viewing angles of up to 30 degrees. The density of the available correspondences obtained using feature points is a key limiting factor in these methods and so ASIFT, a extension to the SIFT feature point that improves performance at wide angles is also introduced. ASIFT is shown to have an order of magnitude increase in correctly matches feature point density at 30 degrees.

Though robust to wide differences in viewing angle, these registration techniques do nonetheless introduce registration errors of up to a few dozen pixels. For this reason the dense SIFT and shifted dense SIFT image comparison algorithms which are robust to registration errors of a few dozen pixels are also developed. The development of these comparison methods includes an analysis of SIFT descriptor statistics and their correlation. Finally these techniques are combined to form an end-to-end change detection system which is evaluated on a number of test datasets.