FAB-MAP: Probabilistic Localization and Mapping in the Space of Appearance

Main idea: SLAM in the space appearance using a probabilistic approach. It is based on the bag-of-words retrieval systems (Níster06). It learns a generative model of the bag-of-words data, i.e. It can identify if a new observation is already included in the map or if it comes from a previously unseen place by using a probabilistic framework.

This approach uses tools as:

• Chow Liu trees which aproximates large discrete distributions. This tree is computed off-line.
  
• A semi-external spanning tree algorithm is used to deal with the size of the Chow Liu tree
  

Probabilistic Navigation using Appearance

• The world is modeled as a set of discrete locations.
• Each location is described by a distribution over appearance words.
• Incoming sensory data is converted into a bag-of-words representation and compared with the stored information.
• There is an expresion for the probability that this observation comes either from the location's distributon or from a place which is not in the map.
  
• This approach uses binary features based on SURF detector/descriptor.

Approximations and Parameters

1. Sets of observations are conditionally independient given position.
2. Detector behavior is idependient of position.
3. Location models are generated independently by the environment.
4. Observations of one feature don't inform about the existence of other features.