Segmentation and Classification for Through-the-Wall Radar Imaging

In this thesis, the problem of stationary target detection, segmentation and classification in Through-the-Wall Radar Imaging (TWRI) is considered. In stationary target detection, Doppler and change-detection-based techniques are inapplicable. A new feature-set that depends on polarimetric signatures and co-occurrence matrices are employed. Algorithms for 2D and 3D segmentation and classification are adapted, investigated and tested. The utilization of these algorithms to the application of TWRI is investigated, with special focus on the feature extraction and target classification phases. A combination of polarimetric signatures and features extracted from co-occurrence matrices is proposed. Two different schemes that deal with 2D and 3D arrangements are proposed. The first scheme is based on a fusion of two dimensional segmentation and classification. The proposed scheme uses features from polarimetric B-scan images to segment and classify the image observations into target, clutter, and noise segments. Target polarimetric signatures from co-polarized and cross-polarized target returns are mapped to a pixel-by-pixel feature space. The image is then over-segmented to homogeneous regions called super-pixels. Homogeneous super-pixels are optionally grouped into clusters and then assigned to corresponding classes. This scheme relies on novel features and the relations between the different pixels. The second scheme deals with the 3D scene directly instead of 2D B-Scans. The proposed scheme uses clustering as an initial phase using intensity and spatial features for each voxel. Clusters that contain mostly noise are ruled out. Further feature extraction using features from the multivariate co-occurrence matrices and polarimetric signatures is applied to the voxels of the remaining cluster(s). Subsequently, the voxels are classified using different classifiers to test the usefulness of the features. This method is designed for practical applications in which the target detection should be performed in real time. The clustering step is used to detect the target positions quickly. Further steps are used to obtain more accurate estimates of the target positions and shapes, and further classify the clustered voxels. All proposed methods are evaluated using real data measurements. The data are collected using three dimensional imaging measurements in a wideband radar imaging scanner exploiting wideband delay-and-sum beamforming