| Synthetic aperture radar (SAR), a high-resolution imaging technique, is unique in remote sensing today in that it permits observation independent of atmospheric conditions and solar illumination. Because the physics of SAR image formation is different from that of conventional optical images at visible wavelengths, algorithms developed for common imaging applications are usually not directly applicable to radar imagery. SAR images are dominated by point scatterers, which can be utilized to extract natural features and recognize man-made objects. This thesis presents an automatic target recognition (ATR) system based on a fundamental analysis of the point scatterers in SAR imagery. First, significant peaks corresponding to objects can be extracted. Then, Delaunay triangulation enables defining texture boundaries of natural features, such as trees and ground. Simulation and experimental results show that texture boundaries can be determined to an accuracy of 2 pixels. Finally, vehicle targets can be recognized by matching the geometrical dimensions of targets with dimensions of vehicle models in the database without exhaustive searching as in other approaches, such as artificial neural networks and template matching. Using synthesized SAR data (Xpatch-generated data), the target recognition algorithm was executed. It shows that a probability of recognition is 50-70%. This ATR system performs reasonably well in the extraction of natural features and the recognition of man-made targets. It can be extended easily when real SAR data and higher resolution SAR data are available in the future. |
