Automatic DEM Extraction from an IKONOS Stereo pair over an Urban Area
Figure 2. Procedure of DEM extraction experiments carried out
The stereo matching algorithm used here (EpiMatch) was developed originally for the SPOT images [Park et al., 2000]. It was based on the findings on epipolarity of linear pushbroom images [Gupta and Hartley, 1997; Kim, 2000; Kim et al., 2001]. This algorithm utilized the knowledge of epipolar curves of linear pushbroom stereo pairs for the estimation of search regions and the determination of patch shapes. A normalized zero-mean cross correlation was used as a similarity measure. It also utilizes an intelligent scheme to determine initial match candidates. It has been reported that the EpiMatch algorithm worked successfully with SPOT images at 10m resolution and KOMPSAT -1 EOC images at 6.5m resolution [Kim and Im, 2001].
The EpiMatch requires the knowledge of epipolar curves. In our case, there can be two methods of estimating epipolar curves of the IKONOS stereo pair. One is using the camera models established based on the DLT model. The fundamental matrix for linear pushbroom images can be derived from the left and right camera models [Gupta and Hartley, 1997]. The other is simply accepting the input IKONOS data in a quasi-epipolar geometry as in a true epipolar geometry. We tried the both method and did not obser ve much difference. In other words, the quasi-epipolar products of IKONOS scenes were in a reasonable accuracy to apply stereo matching.
Stereo matching output was converted to 3D coordinates and these were interpolated to create a DEM. Although sophisticated interpolation schemes or manual post processing can improve the quality of a DEM, a simple Gaussian interpolation was applied. It was not possible to obtain a dense and accurate ground reference DEM over the test area to test the quality of DEM produced in this paper. We instead compared DEM heights with the height of the ground control points used for camera modeling.
