CNN for Radial Velocity and Range Components Estimation of Ground Moving Targets in SAR

Abstract

Ground-moving objects in synthetic aperture radar (SAR) images appear defocused and azimuthally displaced using conventional SAR image formation algorithms. In this paper, a novel regression method based on convolutional neural networks (CNNs) for the estimation of radial velocity and slant range components of ground moving targets is proposed. Motion parameters estimation can be helpful for designing additional matched filters to focus and relocate moving targets. We have generated the training and the test data in such a way that each image is indeed a 2D data matrix of a moving target. In other words, each complex image contains the range-compressed signal of only one moving target with a specified pair of (range, radial velocity). To further decrease the estimation error, we employed transfer learning by fine-tuning the pretrained AlexNet architecture in a regression problem. To verify the effectiveness of the proposed method, simulations have been performed. The results demonstrate the effectiveness of the proposed method.