Deriving accurate GCPs from TerraSAR-X stereo data over urban areas

For autonomous driving, and many other applications, highly accurate reference coordinates are of crucial importance. Many known methods yield a high relative accuracy but lack of absolute coordinates w.r.t. a global reference system. Since metallic objects like lamp poles or traffic signs are common features in urban scenes and appear as focused points in SAR images, their position can be reconstructed from SAR stereo data. This work deals with the detection of such objects in SAR images followed by the reconstruction of their 3D coordinates with high accuracy. For speckle reduction the images are non-locally filtered, thus preserving the input resolution. The Range and Doppler equations are utilized to extract the underlying epipolar geometry and to rectify both images such that the stereo parallaxes are limited to one dimension. In this process a coarse DEM is incorporated to also limit the stereo parallaxes to a minimum. To avoid manual point of interest selection, point shaped objects are extracted by applying a matched filter yielding potential coordinates of such metal poles with pixel accuracy. To get the subpixel center of the objects a local patch is upsampled using FFT and an energy functional is optimized over the gradient magnitude of border pixels. The center of the optimal segment is then determined as the center of mass using the upsampled amplitude values. Potential matching object points are now gathered by taking the nearest neighbor for each point in the second image, where only small parallaxes are allowed. Finally, 3D coordinates are calculated by spatial point intersection with decimeter accuracy. The presented method is very efficient in terms of runtime as image matching is replaced by finding coordinate correspondences. For this reason also ascending and descending pairs can be processed. Best results are achieved by TerraSAR-X highest resolution mode, called Staring Spotlight Mode. In future the sensor models should be adapted by an atmospheric correction.