Métrologie 3D par vision active sur des objets naturels sous-marins

This PhD Thesis concerns the application of computer vision techniques to natural underwater images. Recently, advances in projective geometry have given a strong formalism to computer vision reconstruction algorithms and has allowed real improvements. Nevertheless, many algorithms may have problems with natural scenes. We present here a complete methodology to make a projective reconstruction of natural scenes from underwater images taken with one uncalibrated camera. In the first step, we are interested in extracting robust features which is a necessary step of image processing. Due to the particular scenes we observe (no geometric simple forms, high noise, no knowledge of the environment), we choose a robust implementation of a point detector based on a multi-scale representation of the images. This one allows us to classify points depending on two criteria: robustness against noise and good localization. It is then possible to match these points between images and compute the projective model of the scene with standard robust methods. In our case, we often have really noisy images and standard algorithms cannot be efficient. We propose to apply a preliminary data processing, which used with our multi-scale approach gives good results. Finally, this work has permitted to develop a software for users, experimented or not, to use advanced techniques for computer vision.