Stereo Camera Calibration

[motters]

I've updated the info for stereo camera calibration. This is an
automatic procedure which uses a calibration pattern, and I've tried
to keep it as simple as possible so that anyone could do it. Once I'm
finished writing the calibration program I'll add it to the downloads
section as an executable program.

http://code.google.com/p/sentience/wiki/CameraCalibration

[spig...@gmail.com]

It seems very interesting.
I'm looking forward to seeing more.

SpiglerG

[bohemian]

Regarding the Fish Food, I have the following questions:

1) Does it matter if both cameras are not exactly aligned vertically?
I have a problem with fixing two cameras on the baseline that
makes at least one of the rectified images distorted.

2) In addition, I would like to know what happens if the distance to
the center of the calibration board is greater than 999 mm?

Cheers

[Bob Mottram]

On 28/02/07, bohemian <mzara...@gmail.com> wrote:

Regarding the Fish Food, I have the following questions:

1) Does it matter if both cameras are not exactly aligned vertically?
I have a problem with fixing two cameras on the baseline that
makes at least one of the rectified images distorted.

In general you should try to get the cameras as well aligned as possible.  Good alignment will make things easier for the calibration program and allow better quality range values to be calculated.  However, no matter how well engineered the stereo camera is there will always be some small amount of misalignment which the calibration program calculates as an offset in pixels.  In my case the vertical offset is only one pixel, but the horizontal is offset by about 23 pixels (equivalent to a few degrees difference in the orientation of one camera lens relative to another).

 

2) In addition, I would like to know what happens if the distance to
the center of the calibration board is greater than 999 mm?

The program doesn't care what this value is.  However, if the pattern is too far away then it will become more difficult to locate the centre spot accurately due to pixelation.  In my setup the distance from the cameras to the centre of the pattern along the ground is 500mm.

[bohemian]

Thanks for your help Bob.

It seems that the misalignment of my cameras was unacceptable so I
have decided to totally replace them. I will try to repeat calibrating
the new cameras as soon as I get them.
Actually I don't want to get involved with details of the camera
calibration and I prefer to start dealing with correspondence and
reconstruction problems, though I'm a novice in computer vision.

Cheers

On Feb 28, 9:07 pm, "Bob Mottram" <fuzz...@gmail.com> wrote:

[Bob Mottram]

Fortunately webcams are easy to replace, and I have quite deliberately been using very cheap cameras so that whatever results I get should be obtainable by anyone.

On the other hand there's no dodging the calibration issue if you want to get good results.  Stereo vision involves significant uncertainties, and so you really have to be as methodical as possible to stand any chance at all.

The problem of reconstruction is about one of the hardest in robotics.  There are various approaches which can be taken, which broadly fall into the categories of dense grid mapping and sparse feature tracking.  I've seen one type of mapping system used with some success last year, but even this was rather unsatisfactory since it wasn't capable of mapping and localising simultaneously (SLAM).

One promising approach may be a particle filter based method which updates multiple grid map hypotheses at the same time, similar to DP-SLAM

  http://www.cs.duke.edu/~parr/dpslam/

This might consume a lot of memory, but provided that the computing required scales linearly it should be ok.

On 28/02/07, bohemian <mzara...@gmail.com> wrote:

[bohemian]

The linear runtime is really interesting. Are you working on
localising right now?.

[Bob Mottram]

The localisation and mapping is something which I'm doing at the moment.  At present something like DP-SLAM, with both the robots motion and the mapping process being described by particles, seems like the best way to proceed provided that its computationally tractable.  The best way to do it may be to for a 2D map, which can subsequently be updated with 3D data.  Once the 2D poses of the robot are known then creating 3D grid maps is straightforward.

On 28/02/07, bohemian <mzara...@gmail.com> wrote: