In this paper, we present a solution to the SLAM problem that enables a mobile robot to create visual maps of a given environment. A visual map is formed by a set of significant points that are detected in a three-dimensional space. Those points are extracted from images taken from the environment. In this work, the significant points extracted are used as landmarks in the environment that can be detected from a set of poses in the environment, thus allowing the localization of the robot. We suppose that the robot is equipped with a stereo vision system that enables it to obtain relative measurements to the landmarks while it navigates along the enviroment. In the work presented here, the map is formed by a set of landmarks, each one described by a three-dimensional position in space referred to a global reference system. In addition, each landmark is associated with a visual descriptor that partially differentiates it from others. The solution presented here is based on a Rao-Blackwellised particle filter and allows to simultaneously estimate the path followed by the robot and the most probable map of the environment. The algorithm is validated with results in simulation, supposing that the vehicle possesses a noisy odometry and is able to obtain measurements to the landmarks in the environment, also corrupted with gaussian noise. The main contribution of the paper is the assessment in the parameters of the filter. We propose the most adequate configuration of the filter that enables to solve the SLAM problem. For this purpose we used typical odometry a stereo vision noise models, commonly found in mobile robot applications.