CS771 Artificial Intelligence
Project Proposal
Adam Smyth

Project:
  Development of situation awareness and modeling module for Nomad
   robot.

Building upon:
  Spherical Map Array datatype developed by Hand, Hutchins, and Evelyn.

Key points:
  Fixing problems with existing implementation.
  Creating additional informational functions for accessing data in the
   map.
  Implementing a wall-detecting linear regression system.

Key issues:
  Identification of problems in existing implementation.
  Finding/developing a usable linear regression algorithm.
  Dealing with intersecting walls.
  Moving object detection/filtering.

Current Status:
  Implementation problems will be addressed first. This should, I think,
   be the easiest part.

  I may be able to adapt an existing linear regression algorithm to this
   task. I need to take a closer look at it, and also see about the next
   problem.

  The issue of intersecting walls is that, if I simply regress all
   points that are adjacent to each other, I could end up turning two
   walls into one. Probably the easiest method of checking for this
   would be to do a regression on the 'blob', and if the calculated
   error is above a certain limit, to consider it to be multiple
   objects, and search for new groupings based on the previous
   regression's result. Another option is to look for grouping patterns
   beyond simple adjacency. This later will probably be preferable, in
   the long run, tho it will take more work to develop.

  We won't want to perform regressions on moving objects, so there will
   probably need to be a threshold for number of 'sightings' of an
   object before regression will be attempted on it. The system will
   also need to filter out the data caused by these objects. For this
   reason, perhaps when a new reading is taken in a given direction,
   points BETWEEN this new point and the robot should be DECREMENTED.
   I'll have to build some test data and try this. The problem is that
   doing this will add some processing time to the map update.

Misc notes:
  I'm thinking of maintaining a separate list of derived
   object-representing equations. When an object's position is
   regressed, the resultant data (equation of line, endpoints, error,
   whatnot) is stored in a list to be referenced later. Perhaps as the
   robot scrolls along, and the leading edge of the map is cleared, data
   should be written to the map from this list? So that the robot
   already has an idea that there is a wall ahead, even if it hasn't
   come into sensor contact with it yet.