Having a proper Digital Surface Model (DSM) is crucial in many earth dependent applications such as change detection, 3D urban modelling, urban planning and environmental monitoring. Generating DSM from High resolution stereo satellite images provides good possibilities in this issue. Image matching technique has an important role in DSM generation from stereo satellite images and directly effects on the quality of DSM. Up to now, many image matching algorithms have been proposed, such as Least Square Matching (LSM), Dynamic Programing (DP) and Semi-Global Matching (SGM) which the last one has a higher efficiency than the other methods.

One of the main inputs for SGM algorithm are the epipolar images. Epipolar images are rectified images so that each row at left image corresponds with the same raw at right image (i.e. parallax y is zero). It will cause to accelerate the matching process and reduce the search space from 2D into 1D. Unlike the images with perspective geometry (Frame Camera), the epipolar geometry of linear array images could not modeled with straight line. Therefore, generating the epipolar images from high resolution stereo satellite images has been a research topic in the photogrammetry and remote sensing.

Among the literatures, one can find different epipolar resampling methods proposed in image and object space. All these methods need Rational Polynomial Coefficient (RPC), orientation parameters or ground control points (GCP). Unfortunately, the orientation parameters are not available and RPCs need to be modified. Also measuring the GCPs is not affordable. So it is significant to decrease or completely remove the need for these information in epipolar resampling.

In this paper, we aim to propose a method to build epipolar image by modeling the epipolar curves without any extra information. In the proposed method, dividing the original image into overlapping tiles and utilizing computer vision algorithms to automatically find the corresponding points in stereo images are employed.

Our proposed framework consists of four main steps. The first step is pre-processing of data. In this step, the stereo pair is automatically registered and divided into overlapping tiles. To register images in the absence of RPC, GCPs or other metadata, SURF feature detector and RANSAC algorithm are employed. SURF operator automatically identifies point features in the images. Then RANSAC algorithm filters wrongly detected conjugate points among all points. These matched points will be used to register image pair using an affine transformation.

In the second step, all registered stereo image tiles will be epipolarly resampled. Epipolar geometry for the image tiles of satellite images is equivalent to simple line.  Fundamental matrix and Morgan’s method are used to build epipolar images.

In the third step, disparity maps of corresponding epipolar images are computed using Semi-Global matching (SGM). The SGM algorithm needs large amount of temporary memory for saving matching costs cube and aggregated costs cube. The size of temporary memory depends on the image size and the disparity range. The solution which has been proposed by SGM is to divide the epipolar images into small image tiles. This idea has also followed in our proposed method.

Finally in the fourth step, the result of experiments will be evaluated. The mean and standard deviation of y-parallax for some conjugate points is used to compare the result of different methods.

A stereo-pair acquired by GeoEye-1 high resolution satellite pushbroom sensor is used in our experiments. This image scene includes urban areas over Qom city in Iran. In continue, epipolar images are produced using fundamental matrix and Morgan’s method. The mean and standard deviation of y-parallax is computed for these methods. The results show that our proposed method without using RPCs could produce the epipolar images with sub-pixel accuracy.