Registration of unmanned aircraft systems remote sensing imagery with severe outliers

We present an unmanned aircraft systems (UAS) image registration method to address the inherent severe outliers caused by low image overlap ratios and non-rigid distortions. The method comprises three components to maintain an accurate alignment on overlapping areas while taking advantage of outliers to approximate the non-overlapping areas. First, a penalty matrix is designed to be as the prior from the view of intensity and geometrical discrepancies. Second, a structure constraint is used to directly align the local structures of inliers, and simultaneously pull outliers coherently to reasonable locations. Third, a renewal scheme is designed to organically combine above to form a uniform feature point set registration process, and therein the dynamic SIFT threshold and outlier weight updating are implemented. Experiments on feature matching and image registration are performed using 150 pairs of UAS images and our method outperforms seven well-known methods in most cases.