Synthetic Aperture Sonar Track Registration Using SIFT Image Correspondences

Repeat-pass synthetic aperture sonar (SAS) imaging for change detection and interferometry relies on accurate image registration. SAS image registration is usually performed using a slow correlation-based method, whereas feature-based registration is becoming more prevalent for synthetic aperture radar due to its relative computational efficiency. In this paper we propose a sonar track registration method, from which the estimated track parameters can be used to compute an image registration. The proposed method uses the scale-invariant feature transform (SIFT), a popular algorithm for feature detection and matching, and is demonstrated on a simulated repeat-pass pair of SAS images. A model of the ideal sonar track geometry is presented. A least squares method is formulated for estimating the track registration from a set of noisy feature correspondences, where random sample consensus (RANSAC) is first used to reject outliers. The resulting track parameter estimate corresponds to an image registration aligned to within 0.03 pixels over the scene, which is within the 0.1-pixel resolution recommended for interferometric processing. The ground truth for the simulated ideal data is precisely known, which allowed for the subpixel localization accuracy of SIFT correspondences to be objectively evaluated. The interaction between the RANSAC error threshold and the misregistration error was also evaluated and found to be chaotic. Overall, our results indicate the potential for features to be used for faster coarse registration to reduce the computation time of a more accurate subsequent correlation-based registration.