I am currently trying to use ARToolkit markers  in an indoor
environment for robot localization. I do not have access to the robot
motion model or inertial sensors and hence Kalman filtering is not an
option. I have fiducials that are approx 30cm*30cm in size and the
goal is to localize to cm level accuracy from a large distance.
One information that I do have is the ground truth relative fiducial
transformations and also the transformation of each fiducial wrt the
global origin.
If I just invert the transformation matrices to each camera I get a
very bad location estimate since all errors get magnified.
I am looking for pointers as to how looking at multiple markers in the
same frame and knowing ground truth transformations between them can
be used for achieving more accurate
localization.(Maybe some sort of nonlinear error minimisation like
Gauss-Newton?)

--
Debadeepta Dey,
Research Intern,
Hitech Robotics Systemz,
India.

(Maybe some sort of nonlinear error minimisation like
> Gauss-Newton?)

Yes, multimarker nonlinear optimization improve accuracy and stability a 
lot in my experience. I don't think you can do it using ARToolkit, 
though I may be wrong here - I'm not using ARToolkit myself.
You can get even better  accuracy if some more information about markers 
shape is used (proportions, feature points inside etc) and by using 
specifically designed markers, not just plain squares.
Exact method should depend on requierment, the markers used, camera, 
camera stability, resolution etc. You can use MXRToolkit, which have 
in-built non-linear optimization, or just write your own, latter usually 
better if you have some special requirements. I myself can't use 
MXRToolkit - it's too calculation heavy, and I'm writing code for 
camera-phone.