RE: [Algorithms] Re: Motion tracking from a live video source

[Willem de B. <wd...@mi...>]

This problem is called scene-reconstruction in computer vision, and
there is a lot

of research on it out there. The method you mention coincides with that
of

extracting feature vectors for two frames and use the disparity between
matching

vectors to deduce a (affine) view transformation.=20

=20

Finding the view transform giving a set of matching feature vectors of
images

is fairly straightforward. I would suggest googling for RANSAC and
specifically

SIFT, and Epipolar geometry. A good first source would be

=20

http://www.cs.ubc.ca/~lowe/keypoints/

=20

Extracting feature vectors that are partially invariant to changes in
illumination,

as well as affine transformations is quite hard. SIFT seems to be very
good

at it, but I'm not sure if this is feasible in real-time.

=20

Cheers,

=20

---

Willem H. de Boer

Homepage: http://www.whdeboer.com <http://www.whdeboer.com/>=20

=20

=20

________________________________

From: gda...@li...
[mailto:gda...@li...] On Behalf Of Ken
Noland
Sent: 11 October 2005 20:28
To: gda...@li...
Subject: Re: [Algorithms] Re: Motion tracking from a live video source

=20

Actually, funny you ask this because I've been debating on writing a
small demo that uses the camera as an input to orient the perspective in
a 3D game. Kinda like a big trackball using the camera to point and
shoot.=20

=20

The idea I was toying around with and starting to seriously consider was
to analyze the frame to determine a grid of distinctive points, say
16x16 distinctive points with color gradient stored, and possibly even
some edge detection in there to sample the most distinctive points on
the 2D image. The next frame it would search for the 16x16 points and
attempt to match up the difference with the current frame and then that
*should* spit out a 3D coordinate change.=20

=20

However, there are certain things that would disrupt that really basic
algorithm. The first of which is moving objects in the scene. That is
why I decided to have a distinctive weighting value added to each point
chosen to sample and each frame that goes by that contains that
particular point at a high degree of accuracy then the weighting value
is incremented, otherwise the weighting value is decremented.=20

=20

Has anyone else experimented with this?

=20

-Krad

=20

On 10/11/05, Juhana Sadeharju <ko...@ni...> wrote:=20

>From: Aras Pranckevicius <ne...@gm...>
>
>PiAgQXQgdGhlIG1vbWVudCwgSSdtIHRyeWluZyB0byBmaWd1cmUgb3V0IGEgY291cGxlIG9
mIHRo=20
>aW5ncyByZWxhdGluZyB0bwo+IHdlYmNhbSBtb3Rpb24gdHJhY2tpbmcuCgpJZiB5b3UgY2F
uIHNv

[ Reposting the coded message in plain ascii. Juhana ]

>  At the moment, I'm trying to figure out a couple of things relating
to=20
> webcam motion tracking.

If you can somehow segment "foreground" (moving object) and
"background" (the rest), then you've got a black&white image. From
there, quite many approaches exist; quite easy is MHI (motion history=20
images) - basically, just fade out the foreground along time, so you
have fading out "motion trails". Then compute image gradients of that.

Of course, all this assumes you've got background extraction working,=20
which is hard :)

Anyway, I suggest taking a look at OpenCV library - it has lots of
vision-related algorithms already implemented:
http://sourceforge.net/projects/opencvlibrary It won't solve your
problem immediately, but at least can provide some building blocks.

--
Aras 'NeARAZ' Pranckevicius
http://nesnausk.org/nearaz | http://nearaz.blogspot.copad: fix the code



-------------------------------------------------------
This SF.Net email is sponsored by:
Power Architecture Resource Center: Free content, downloads,
discussions,=20
and more. http://solutions.newsforge.com/ibmarch.tmpl
_______________________________________________
GDAlgorithms-list mailing list
GDA...@li...
https://lists.sourceforge.net/lists/listinfo/gdalgorithms-list
Archives:
http://sourceforge.net/mailarchive/forum.php?forum_id=3D6188

=20