This thesis aims at real time tracking of 3D objects, in order to develop augmented reality applications. Augmented reality needs a stable and accurate tracking. Adding that we also want real time video tracking, we had to found compromises between results's precision and speed of treatment. This report contains the description of three tracking algorithms developed during this thesis. They illustrate the progression accomplished by our work during these three years, which is the tracking of increasingly complex objects, first planar objects, then simple 3D objects, and finally complex 3D objects with an unknown model. With the first algorithm we were able to track few textured planar objects. It is an extension of a fast and accurate algorithm that tracks textured plans. We added a contour component to this algorithm to make it track a greater amount of patterns. When this work on planar tracking have been done, we adapted the texture tracking algorithm to 3D objects. Using multiple occurrences of this algorithm distributed on the object's surface to follow, and associating them with an iterative algorithm of pose estimation, we managed to track in real time some simple objects during their translations and rotations up to 360 degrees. This algorithm was limited by the fact that a 3D model of the followed object had to be known. That's why we tried to find an algorithm that would be able, during an off-line learning stage, to generate a statistical model of the object from key-views. Based on the same texture algorithm that the one used in previous algorithms, it doesn't estimate the objects' pose, but characterizes it by a 2D spline deformation.
