Invariant models for ground control points in high resolution images

This paper addresses the problem of automatic identification of ground control points (GCPs) in high resolution satellite digital images. The method is based on characteristics of GCPs and their neighbourhood, invariant with respect to rotation and contrast change. The characterization algorithm, which leads to a feature space with a fixed number of dimensions a critical advantage of the method is based on the autocorrelation function of an azimuthal projection around the GCP. Such features are first extracted from a set of reference images and stored in a database, together with the GCP geographical co-ordinates, and used in Euclidean minimum distance classifier, looking for matches with pixels of the image being geo-referenced. Practical issues lead to the necessity to filter out potential wrong identifications, through a consistency filter based on geometrical patterns made of ambiguous GCPs. The performance of the method is studied, demonstrating its insensitivity with respect to rotation and variations of contrast, its discrimination power, and its sensitivity to determine the absolute position of a GCP to within a fraction of pixel.