Automatic DEM Extraction from an IKONOS Stereo pair over an Urban Area
IKONOS images over Taejon metropolitan city of the Republic of Korea were used for the experiments. The left and right pair was acquired on the same day of 19
th November, 2001. They were precision-geometrically corrected and then resampled to a quasi-epipolar geometry (provided by the courtesy of e-HD. com). The size of each image was 9700 by 9616 pixels and the images were believed to be subtracted from original full-size images. The images cover approximately 8kms by 8kms on the ground. Figure 1 shows the IKONOS image (left) . The image contains dense population of residential houses, apartments and industrial buildings as well as rivers and hills. Automated and reliable DEM extraction from such images is a challenge to any stereo matching algorithms.
Figure 1. An IKONOS image over Taejon metropolitan city (Left image only). The points with crossed circles were used as control points used for camera modeling and tie points for stereo matching.
2. DEM Extraction Method
This section briefly explains the DEM extraction method used in this paper. The procedure of DEM extraction is shown in figure 2. We first establish camera models for the left and right images. A suitable camera model for IKON OS images is required in this step. In our experiments, we chose the Direct Linear Transformation (DLT) model proposed by Gupta and Hartley [1997] as a camera model. Note that the IKONOS images provided were already geometrically corrected and resampled to a quasi-epipolar geometry. We decided to ignore the transformations previously applied to the dataset and regard the geo-product as raw images, instead of trying to find out all transformations between the original raw images and the geo-product provided and to decide a suitable model given these transformations. Since the DLT is a generalized camera model and does not require any sensor-specific knowledge, camera models of the left and right images could be established. Also it is notable that the DLT is the simplest case of
rational function models. Using ground control points obtained from GPS surveying, the left and right camera models were established. These camera models were then used for later processing.
