Scene Features in Vision-SLAM

A.J Davison's map-building and localization system is based on discrete features, like the majority of robot navigation systems. Discrete features are easier to represent than continuous objects or regions, and provide unambiguous information for localization. The spatial density of features used can dier depending on how well it is required to represent the world, and individual ones can easily be added or deleted as necessary.

The features which will make up the sparse map used for localization in this work need to be "landmarks" for the robot: they must be reliably and repeatedly detectable and measurable, and preferably from a wide range of robot positions. This differs slightly from the requirements of the features used in structure from motion algorithms where it is not usually intended that the same features (usually "corners" or line segments) be detected and matched over long periods | indeed this is typically not possible because with a passive camera features can go out of view very quickly. What properties do landmarks need to have?

• They must be features present in a normal environment so that no articial interference such as positioning of beacons is required.
• They must be stationary in the scene (or potentially have known motions, in work
  beyond the scope of this thesis).
• They must be easily identifiable from a variety of ranges and angles, and not frequently
  be occluded.

It was decided to restrict landmarks to being stationary, point features, and to use simple image patches to represent them, with matching to be performed using normalized sum-of-squared-difference correlation. The patches used are larger than those usually representing corner features in structure from motion systems: 15 *15 pixels rather than 3*3 or 5*5. The reason for this choice is that with larger patches the chances of mismatches are reduced due to their greater distinguishability. Having larger patches makes processing slower than otherwise, but the relatively small number of features used and the fact that with the active approach only one feature is observed at a time means that this is not a problem. The key to our approach is to have a small number of very reliable landmarks rather than a large number of transient ones.

Keyword: SLAM Corner detector

Cited from A.J Davison, OX "Mobile Robot Navigation Using Active Vision"