Relative orientation and roto-rectification of stereoscopic underwater images

This work has been developed under the scope of a stereo vision project for autonomous underwater vehicles (so called AUV). Our solution consists in a vehicle provided with two optical sensors, emulating the binocular system of human beings. In order to obtain better precision for the numerical model of a subaquatic scene, calculus are made from pairs of images before and after vehicle removal. This information, makes the searching easier, allowing modelization of certain stereoscopic perceptual response aspects when introducing a reference numerical model previously stored. Starting from an automatic identification of a reduced set of homological points on images (subject of another working line inside this stereo project), our proposal in this work consists, in the first place, in improving the image rotations (derived themselves from navigating vehicle system) using relative information from identified homological points. Then, locating which plane is optimal for projecting images, and finally, recalculating pixel values while projecting (becoming a roto-rectification for both images). Once reaching two roto-rectified images, we would be able to calculate the numerical model, with good performance algorithms and enough precision. These last algorithms, that represent a third searching line inside this project, have also been the subject of paper presentations.