g., [2�C6]). Typically, these applications use directly the laser measurements to build 2D occupancy grid maps [2, 5], selleck catalog or Inhibitors,Modulators,Libraries they extract features from the laser measurements [4, 7] to build 2D landmark-based maps. Nevertheless, recently the interest on using cameras as sensors in SLAM has increased and researchers focus on the creation of three dimensional maps based on the measurements provided by vision sensors. These approaches are usually denoted as visual SLAM. Compared to laser ranging systems, stereo vision systems are typically less expensive. In addition, typical laser range systems allow to collect distance measurements on a 2D plane, whereas the information provided by stereo vision systems can be processed to provide a more complete 3D representation of the space.

On the contrary, stereo systems are usually less precise than laser sensors. In common configurations, the camera is installed at a fixed height and orientation with respect to the robot Inhibitors,Modulators,Libraries reference system and the movement of the camera and robot is restricted to a plane [8, 9].The research in visual SLAM has many similarities with the rich research in the Structure Inhibitors,Modulators,Libraries from Motion (SFM) field, carried out in the computer vision community. The approach taken in SFM, however, has generally been very different from visual SLAM solutions because the applications did not require real-time operation, and the trajectory of the camera and the structure of the environment could be computed offline. Some real-time SFM systems have been produced by efficient implementation of frame-to-frame SFM steps (e.g.

, [10]), in which repeatable localisation is possible and motion uncertainty does not grow without bound over time. Visual SLAM approaches typically deal with large camera trajectories and Inhibitors,Modulators,Libraries significant Cilengitide uncertainties in order to compute a visual map online.Stereo vision systems provide a huge quantity of raw information from the environment stored in both images. In consequence, the images are normally processed in order to reduce the information to be used for mapping. As a result, most approaches to visual SLAM are feature-based. In this case, a set of points extracted from the images are used as visual landmarks. Features, such as image edges were used in [11] to build maps using a single camera. However, the localization of the camera with respect to the segments is difficult.

For example, KOS 953 in [12] regions of interest are extracted using a visual attention system. The regions are extracted from images at different scales in a similar manner as the human perceptual system does. The main drawback with this kind of landmarks is that it may be difficult to obtain an accurate measurement of a region using a stereo camera, since regions can be arbitrarily large, thus providing inaccurate results.In this paper we extract salient points from images and use them as visual landmarks in the environment.