Surface reconstruction from interferometric ISAR data

Abstract

A frequently useful technique in the interpretation of Inverse Synthetic Aperture Radar (ISAR) images is to construct a three dimensional (3D) model of the object being imaged. Generally, such models are constructed manually by an analyst based on a series of radar images and whatever other information is available. However, using multistatic radar, it is possible to generate 3D Interferometric ISAR (IFSAR) point cloud images. In this thesis, two original techniques for automatically generating models of rigid bodies from IFSAR point clouds are explored. One technique extends the concept of the visual hull to a composite point cloud. The other uses a noise resistant estimator to determine the shape of the side of the object presented to the radar. Noise and radar artifacts show up strongly in the data, and both techniques must reject them to achieve good performance. Additionally, an optimization-based algorithm was devised to determine the angular velocity of a target using only one interferometric baseline. Knowing the angular velocity of the target is necessary to correctly scale the axes of ISAR images and IFSAR point clouds. Though techniques exist for angular velocity determination using multiple baselines, receiving antennas are expensive and are not always available. The techniques presented in this thesis were tested against simulated data and data collected in a compact range. The angular velocity determination technique was successfully demonstrated on simulated IFSAR data, using a particular heuristic enforcing the consistency of rotational motion. Investigation into a more robust heuristic is necessary to make the approach broadly effective. The surface reconstruction algorithm based on the noise resistant estimator performs very well, doing much better than a traditional algorithm selected for comparison (3D a-shapes) in high noise situations. The technique based on the visual hull, while producing faithful reconstructions in some cases, generally offers performance inferior to the noise resistant estimator. Quantitative measurements were used to evaluate the fidelity of the models generated by the various techniques.