Abstract: Towards the issue of epipolar resampling for linear array satellite stereo-imagery, we analyze the advantages and disadvantages of the off-the-shelf methods. Then a new epipolar resampling approach that combines the Object-based and Image-based method is proposed. The approach has two steps. Firstly, the left epipolar image is resampled with the Image-based technique. Specifically, we divide the left original image space into several regular rectangular grid, then each grid is rotated independently to make epipolar direction parallel to the image line. After that resampling is achieved using local affine transform model. Secondly, both the right and left image are projected onto a datum plane in the object space to resample right epipolar image that has the same resolution as the left image. Experiments on five pairs of stereo pairs and comparison with the two prevailing methods show that the proposed method not only can obtain the epipolar image product with vertical parallax close to zero and superior to the comparative method, but also can guarantee the linear proportional relationship between horizontal parallax and elevation. In particular, the experimental results of three groups of larger epipolar line curvature data confirm that this method has good robustness and applicability, the proposed method could be well applied to different sensor data processing systems.

Key words: linear array satellite, local affine, epipolar image, RPC, vertical parallax