| |
 |
|
|
Science Forum Index » Image Processing Forum » Problem with object registration...
Page 1 of 1
|
| Author |
Message |
| ... |
 Posted: Thu May 08, 2008 1:32 am |
|
|
|
Guest
|
Hi,
i have printed a squared marker on a paper and i use it in order to do
augmented reality.
I have the following problem. When i found the marker in the image, i
need to find the four corners in order to compute the camera extrinsic
parameters and render a 3d object (a cube) on the marker (see for
example artoolkit).
The problem is that sometimes it does not work well. I mean, if the
position of one of the corners is changed of only one unit (for
example, from (160,180) to (161,180)) there is a noticiable difference
in the object rendered on the marker.
There is an even bigger problem. If, near the marker, i put some other
extra stuff (for example, a code that identifiy the marker), than with
the matrix computed from the four corners of the marker (the one that
transform a world point to a frame buffer point), i would like to
extract what is near the marker, in order to have the id of the
marker. But a smal change in the coordinates of the corners imply that
the id is not extracted well....for a "small change" i mean the fact
that the contours of the marker is not perfectly defined (this problem
is due to the camera, the resolution of the image, which is 320x240
etc...)...there is no way to correct the positions of the four
corners...
The photo of the marker was really clear, and the distance from the
camera to the marker was about 30 centimeters (or less). The bounding
box of the marker on the 320x240 image was 37x24.
Maybe the marker is too small (or equivalently i was too distant from
the marker with the camera)??
I don't know how to fix this....help!!
Thank you |
|
|
| Back to top |
|
| Jonathan Campbell... |
| Posted: Thu May 08, 2008 8:48 am |
|
|
|
Guest
|
kingpin at (no spam) freemail.it wrote:
Quote: Hi,
i have printed a squared marker on a paper and i use it in order to do
augmented reality.
I have the following problem. When i found the marker in the image, i
need to find the four corners in order to compute the camera extrinsic
parameters
Please enumerate the extrinsic parameters. Please also tell us what
algorithm you used and what input data (coordinates of four corners?).
I have no particular expertise in camera calibration, but I'm thinking
'control points' and image affine transformation (from one coordinate
system to another) --- as one does when transforming satellite images
from image coordinates to map coordinates.
In that case, essentially 2D -> 2D, we have six parameters, the four
components of the linear transformation matrix and the two translations.
In the case of the four matrix components, we may have just three
independent parameters (rotation angle about the optical axis and tow
scale parameters).
You probably have nine + three parameters (3D -> 3D). But maybe, for
example assuming image plane parallel to object (square) plane and
camera axis perpendicular to both those, you can make the problem 2D ->
2D. As above the nine matrix parameters are not truly independent.
In any case, thinking back to the 2D geometric calibration problem, we
always liked to have 16 or 20 control points distributed reasonably
evenly across the image. A small number of points close together would
guarantee an ill-conditioned system of equations --- which would result
in behaviour such as you mention.
Incidentally, it was never considered feasible to use the minimal set of
equations; always overdetermined.
If you decide to use multiple fiducial marks, e.g. one at each corner of
your object plane (or more), then crosses or other shapes may be better
than squares. Google <fiducial mark>.
Best regards,
Jon C. |
|
|
| Back to top |
|
| |
|
Page 1 of 1
All times are GMT - 5 Hours
The time now is Mon Sep 08, 2008 1:23 pm
|
|