On 18 Lug, 07:14, Yoong Kang <yoongkang....@[EMAIL PROTECTED]
> wrote:
> Hi all.
>
> I am new to computer vision. Currently, I am doing a stereo vision
> project. As I understand it, the two main stereo vision methods are
> feature-based and area-based. I am required to use a feature-based
> method, because I am restricted to a very specific edge-detection
> algorithm (http://www.tfolsom.com/Research/Research.htm).
This edge-
> detection algorithm returns information about the edges in a text file
> (i.e, position, orientation, etc.).
>
> Since the edge-detection part has already been done, the next step is
> stereo matching. I have tried looking at a number of research papers
> on the topic, but most of the information I was able to find was on
> dense methods (non-feature-based).
>
> I still do not know how a feature in a left image is matched with the
> correct feature in the right image (and not other features). What is
> the criterion that is used to match the features?
>
> Would anyone be kind enough to point me in the right direction? I am
> currently confused as to how to begin this project.
>
> Thanks in advance.
A 'feature' is not just a point in an image but should have some
attributes. As an instance, if it is a vertex, it has an orientation
which, not considering projective distorsion, should be comparable
between the two images.
Things are also different if you are using rectified or non rectified
images. If your cameras are calibrated, you are able to draw pencils
of corresponding epipolar lines on the two images. This simplifies
your task, because features on a left epipolar line can only match
features on the corresponding right epipolar line. Then you have a
plethora of methods for matching them, some of which are reviewed in
chapter 6 of Faugeras, Three-Dimensional Computer Vision - A Geometric
Viewpoint (1993: not really up to date, but it can be a starting
point).
If your cameras are uncalibrated, several other methods can be used to
match features. See eg this paper:
www.hpl.hp.com/techre****ts/97/HPL-97-96.pdf
describing a revisitation of the old Longuet-Higgins method.
|
