This thesis presents the realization of a localization system for a mobile robot relying on monocular vision. The aim of this project is to be able to make a robot follow a path in autonomous navigation in an urban environment. First, the robot is driven manually. During this learning step, the on board camera records a video sequence. After an off-line processing step, an image taken with the same hardware allows to compute the pose of the robot in real-time. This localization can be used to control the robot and make it follow the same trajectory as in the learning step. To do this, we build a three dimensional model of the environment (a map) and use this map to localize the robot. The map building step is a structure from motion algorithm. The map is a three dimensional model of the environment made of points. In order to localize the robot, we establish correspondences between the curent image and the 3D model. This allows to compute the pose of the camera and the robot. Even if the goal is to be able to localize a vehicle in an urban environment, we have chosen not to make restrictive assumptions on the structure of the scenes. Additionally, the pose computation is done with six degrees of freedom which doesn't limit the usage of the system to vehicles moving on a ground plane. Finally, a large part of this report is dedicated to the performance evaluation of the system. Many experiments in real situations were conducted to evaluate the robustness. The accuracy was measured by comparing the localization computed with the vision algorithm to the ground truth recorded with a differential GPS sensor.