Estimating the motion of a mobile robot while simultaneously building a representation of its environment is a key problem for autonomous robotics. This problem is known as SLAM (Simultaneous Localization and Mapping). In this thesis, we address the problem of Bearing-only Visual SLAM based on an omnidirectional camera. The environment is made of point-based landmarks whose depth is never measured directly. We first made a state of the art of SLAM solving methods. We finally kept two algorithms because of their good consistency: the SAM (Smoothing and Mapping) which is a probabilistic method based on the smoothing of the whole robot trajectory, and the interval analysis. Simulation results shown that the interval method is not well adapted to the bearing-only SLAM: the system is not redundant enough. The SAM algorithm was then validated on real data, with two planar trajectories and a full six degrees of freedom trajectory. Complexity is a classical issue of SLAM methods: computational cost quadratically rises with the number of landmarks. This issue is often solved by splitting the whole environment in local submaps. We used this concept and improved it on two points. First, each new local map consistently takes advantages of all the measurements made in the previous one. Then, we developed a new criterion to decide when starting a new map which ensures that the current map remains strongly correlated. This leads to limit the growth of uncertainties in local maps. Finally, we used the latter property to apply deterministic algorithms to enhance the point-based representation by textured planes and a pertinent representation of the free space.
