Inverse stereo image matching for change detection

摘要

A system and method for finding real terrain matches in a stereo image pair is presented. A method for finding differences of underlying terrain between a first stereo image and a second stereo image includes performing epipolar rectification on a stereo image pair to produce rectified image data. The method performs a hybrid stereo image matching on the rectified image data to produce image matching data. A digital surface model (DSM) is generated based on the image matching data. Next, the method identifies areas in the DSM where the stereo image matching should fail based on the image matching data and the DSM to generate predicted failures. The method can then determine real terrain changes based on the predicted failures and the image matching data.

主权项

1. A method for finding differences of underlying terrain between a first stereo image and a second stereo image comprising:
performing an epipolar rectification on at least one of the first stereo image and the second stereo image to produce rectified image data; performing stereo image matching on the rectified image data to produce image matching data; generating a digital surface model (DSM) based, at least in part, on the image matching data; identifying areas in the DSM where stereo image matching fails based, at least in part, on the image matching data and the DSM; among the areas in the DSM where stereo image matching fails, identifying predicted occlusion failure areas in which the stereo image matching should fail due to occlusions, wherein the predicted occlusion failure areas do not include real terrain changes; identifying remaining areas among the areas in the DSM where stereo image matching fails as unpredicted failure areas that include real terrain changes, wherein real terrain changes include at least one of the group of: a first three-dimensional object moving between two different stereo images, a second three-dimensional object appearing in the second stereo image and not in the first stereo image, and a third three-dimensional object appearing in the first stereo image and not in the second stereo image; separating the predicted occlusion failure areas from the unpredicted failure areas to determine the real terrain changes; and generating a new image based on the real terrain changes, wherein at least some of the real terrain changes are eliminated from the new image, and wherein the method is performed without performing ortho-rectification on the first and second stereo images.