EM estimation of a generic 2D object model based on a sparse set of incomplete ISAR images

This paper addresses the estimation of a two-dimensional model of an object, based on measurements with a network of High Range Resolution (HRR) scanning surveillance radars. While considering a dynamic radar scene, the data collected from the multiple radars at multiple scans of the antenna provide a wide, but highly sparse, coverage in 2D space. The multi-radar multi-scan data are treated as asynchronous. Inverse Synthetic Aperture Radar (ISAR) 2D imaging is independently performed within the narrow angular sector which is covered during each single radar scan. Each of the formed ISAR images is a rotated and incomplete version of the 2D profile of an extended object. The multi-look images are aligned and incoherently fused in a common 2D space. Their complementary information content is thus combined in a unique composite image. The Expectation- Maximization (EM) algorithm is further applied for parametric estimation of a low-dimensional 2D object model. This model constitutes a feature vector, useful for object classification. The local-only use of the phase in this overall sparse, but locally dense, sampling scheme accelerates significantly the de-ghosting of the formed 2D target images. Quicker convergence to an unambiguous estimate of the generic 2D object model is achieved, as compared to a purely incoherent processing of the sparse dataset.