This thesis presents a nonlinear method to estimate the fundamental matrix, a key problem arising in projective motion estimation and reconstruction, based on a general Errors-in-Variables (EV) model. In this model, the method considers that all the measurements are corrupted by noises, and minimizes a cost function derived from a nonlinear criterion to estimate both the fundamental matrix and corrupted data, involving the rank-2 constraint With reasonably adjusted data, this method turns out to significantly increase the accuracy and robustness with a simple form of computation. The performance of the proposed approach is justified by theory and assessed by several experiments on real images.