Analysis of methods

We can use as base the approach at GRASP lab:

• It uses a geometric approach (epipolar geometry) 5 and 3-point algorithms combined with a Preemptive RANSAC to compute accurately the position of the camera.
• It also reconstruct the scene from 3D-2D correspondences.
  
• Based on that is able to perform localization and mapping.
• Loop-closing is performed using a bag-of-words-based approach.
• GPS data is used to restrict the search area. (This is not possible in indoors).
• It also uses epipolar geometry to avoid wrong matches.

To apply this approach in a topological layer we should:

• Define the nodes in the topological map. These can be represented as:
• Fingerprints (Tapus 2005)
• Epitomes (Ni 2008)
  
• unsupervised learning approach used to create clusters based on the information present in the acquired images. (Kuipers 2002, Bowling 2006)
  
• Define the relation between these nodes:
• Metric information (distance, angle)
  
• Conectivity to the neighbors.
  
• Define the actions to be performed by the robot. These actions can be:
• Explore new space
• Go to the center of free space
• Leave room
• Follow the mid-line
• Etc

Some works that use a topological approach to perform appearance-based SLAM are the following:

• Appearance-Based Topological Bayesian Inference for Loop-Closing Detection in a Cross-Country Environment (Chen and Wang 2006).
  
• They use PCA to model the environment's appearance.
• Their distributions are approximated by a series of Gaussian models.
• FAB-MAP: Probabilistic Localization and Mapping in the Space of Appearance
• It is based on the bag-of-words retrieval systems (Níster06).
• 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 (loop-closing).
• It uses SIFT/SURF detector/descriptor.
• 
  
• See the references in the blog.

Proposal:

• FAB-MAP can be extended adding metrical information provided by epipolar geometry and scene reconstruction.