Computer vision in AUVs: automatic roto-rectification of stereo images

This work has been developed in the framework of a project of stereo vision for autonomous underwater vehicles (AUVs) provided with optical sensors. To obtain a numerical model of an underwater scene of more precision than the one that would be obtained starting from a couple of simultaneous images, we should re-calculate starting from two images taken in consecutive way. The increase of the distance among the points of view would improve this precision. Normally, in consecutive exposures, the lines of bitmaps do not coincide with the epipolar lines as in case of simultaneous images from calibrated sensors. This complicates the search of homologous points. Parameters of relative rotation and translation of two consecutive images are approximately known. To know them with more accuracy allows us to carry out the image roto-rectification and to use the results. Obtaining knowledge of a certain number of couples of homologous points, we can calculate with enough precision the relative rotation parameters in order to proceed with image roto-rectification, obtaining a new image couple where epipolar lines match with bitmap lines. This involves three major steps. The first one concerns the obtention of relative orientation among components of a stereo couple. This relative orientation is then used to find an optimal plane to minimize deformation when projecting images. Finally, the projection itself of the couple over an optimal plane is found. Then, it is possible to apply the 3D reconstruction process. In conclusion, given any stereo pair of images, the whole process of 3-D numeric model obtention can be automated