23 May 2006
Arnau Ramisa

The ability to navigate, namely go from any starting point to any destination avoiding obstacles, is essential for any robot that wishes to be called autonomous. This difficult task can be divided in several strongly related sub-tasks: map building, which attempts to build a useful representation of the environment; localization, that establish the position of the robot in some map; path-planning and obstacle avoidance. In principle, the robot odometry should be enough to solve the localization problem, but it is well known that odometry has an additive error, which turns useless localization information during a long drive. To solve this problem the Simultaneous Localization and Mapping techniques where developed. This techniques iteratively build a map and use it to correct the localization errors. We propose a topological SLAM system for indoor environments that describes every room with one or various panoramas of affine covariant feature points, extracted from the images captured by a camera. Each panorama will the be linked to a graph where edges represent adjacency between rooms. This approach is interesting because it eases the path-planning tasks while allowing, to a certain degree, the extraction of metric information.