Image Registration and Augmented Reality

classic Classic list List threaded Threaded
6 messages Options
Reply | Threaded
Open this post in threaded view
|

Image Registration and Augmented Reality

luca.dantona
Hi!
Is anybody aware of the existance of some sort of documentation on
this topic?
I'm trying to detect the transformation matrix between consecutive
frames in a video with feature points detection (SIFT) and calculating
the optical flow with the Lucas-Kanade method in pyramids.
Therefore I used RANSAC to estimate the best omography transormation
and added a layer of augmented reality with cvWarpPerspective.
The main two techniques I found are the following:
- FrameToReference: I get the feature points from the first frame
(reference frame) and then compute the matches in the follwing frames
with the first as reference. The problem here is that after a while
the current frame could be very different from the reference one, and
the matching of the correspondences too hard.
- Frame to Frame: The matching of feature points is always calculated
between consecutive frames. This could be simpler, because consecutive
frames are very similar, but on the other hand, the homography is
incremental and every little error in each step adds up; after a few
seconds the omography is too wrong.

Is there a way to mix up this two techniques to get better results?
Thanks :)

Reply | Threaded
Open this post in threaded view
|

Re: Image Registration and Augmented Reality

khos88
Hi

I also study image registration based KLT Tracker. But this tracker fails when features are tracked between frames with large motion. and also there are many lost features between frames.

please give me advise for imroveing KLT Tracker

best regards
Sukhee


--- On Wed, 1/14/09, luca.dantona <[hidden email]> wrote:
From: luca.dantona <[hidden email]>
Subject: [OpenCV] Image Registration and Augmented Reality
To: [hidden email]
Date: Wednesday, January 14, 2009, 4:14 AM










   
            Hi!

Is anybody aware of the existance of some sort of documentation on

this topic?

I'm trying to detect the transformation matrix between consecutive

frames in a video with feature points detection (SIFT) and calculating

the optical flow with the Lucas-Kanade method in pyramids.

Therefore I used RANSAC to estimate the best omography transormation

and added a layer of augmented reality with cvWarpPerspective.

The main two techniques I found are the following:

- FrameToReference: I get the feature points from the first frame

(reference frame) and then compute the matches in the follwing frames

with the first as reference. The problem here is that after a while

the current frame could be very different from the reference one, and

the matching of the correspondences too hard.

- Frame to Frame: The matching of feature points is always calculated

between consecutive frames. This could be simpler, because consecutive

frames are very similar, but on the other hand, the homography is

incremental and every little error in each step adds up; after a few

seconds the omography is too wrong.



Is there a way to mix up this two techniques to get better results?

Thanks :)




     

   
   
       
         
       
       








       


       
       


     

[Non-text portions of this message have been removed]

Reply | Threaded
Open this post in threaded view
|

Re: Image Registration and Augmented Reality

Juan D'Amato
Do you know AR Toolkit ?

It's a SDK for building Augmented Reality Apps

I recommed you to check it out

http://www.hitl.washington.edu/artoolkit/

Good luck

Juan

--
Ing. Juan P.D'Amato
Tel : 02293 - 439690
Web Site : www.pladema.net/~jpdamato/


[Non-text portions of this message have been removed]

Reply | Threaded
Open this post in threaded view
|

Re: Image Registration and Augmented Reality

abhijit kundu
In reply to this post by khos88
@Sukhee

There was a publication in CVPR 2008 by Stan Birchfield, which
improved upon the the KLT. See
here:http://www.ces.clemson.edu/~stb/research/jointtracking/


regards,
Abhijit

On Fri, Jan 16, 2009 at 11:23 AM, Sukhee Ochirbat <[hidden email]> wrote:

> Hi
>
> I also study image registration based KLT Tracker. But this tracker fails
> when features are tracked between frames with large motion. and also there
> are many lost features between frames.
>
> please give me advise for imroveing KLT Tracker
>
> best regards
> Sukhee
>
> --- On Wed, 1/14/09, luca.dantona <[hidden email]> wrote:
> From: luca.dantona <[hidden email]>
> Subject: [OpenCV] Image Registration and Augmented Reality
> To: [hidden email]
> Date: Wednesday, January 14, 2009, 4:14 AM
>
> Hi!
>
> Is anybody aware of the existance of some sort of documentation on
>
> this topic?
>
> I'm trying to detect the transformation matrix between consecutive
>
> frames in a video with feature points detection (SIFT) and calculating
>
> the optical flow with the Lucas-Kanade method in pyramids.
>
> Therefore I used RANSAC to estimate the best omography transormation
>
> and added a layer of augmented reality with cvWarpPerspective.
>
> The main two techniques I found are the following:
>
> - FrameToReference: I get the feature points from the first frame
>
> (reference frame) and then compute the matches in the follwing frames
>
> with the first as reference. The problem here is that after a while
>
> the current frame could be very different from the reference one, and
>
> the matching of the correspondences too hard.
>
> - Frame to Frame: The matching of feature points is always calculated
>
> between consecutive frames. This could be simpler, because consecutive
>
> frames are very similar, but on the other hand, the homography is
>
> incremental and every little error in each step adds up; after a few
>
> seconds the omography is too wrong.
>
> Is there a way to mix up this two techniques to get better results?
>
> Thanks :)
>
>
>
>
>
>
>
>
>
>
>
> [Non-text portions of this message have been removed]
>
>
Reply | Threaded
Open this post in threaded view
|

Re: Image Registration and Augmented Reality

khos88
Dear Sir

Thank you for your reply

good luck for you
sukhee

--- On Fri, 1/16/09, abhijit kundu <[hidden email]> wrote:
From: abhijit kundu <[hidden email]>
Subject: Re: [OpenCV] Image Registration and Augmented Reality
To: [hidden email]
Date: Friday, January 16, 2009, 10:51 AM










   
            @Sukhee



There was a publication in CVPR 2008 by Stan Birchfield, which

improved upon the the KLT. See

here:http://www.ces. clemson.edu/ ~stb/research/ jointtracking/



regards,

Abhijit



On Fri, Jan 16, 2009 at 11:23 AM, Sukhee Ochirbat <khos88@yahoo. com> wrote:

> Hi

>

> I also study image registration based KLT Tracker. But this tracker fails

> when features are tracked between frames with large motion. and also there

> are many lost features between frames.

>

> please give me advise for imroveing KLT Tracker

>

> best regards

> Sukhee

>

> --- On Wed, 1/14/09, luca.dantona <luca.dantona@ yahoo.com> wrote:

> From: luca.dantona <luca.dantona@ yahoo.com>

> Subject: [OpenCV] Image Registration and Augmented Reality

> To: OpenCV@yahoogroups. com

> Date: Wednesday, January 14, 2009, 4:14 AM

>

> Hi!

>

> Is anybody aware of the existance of some sort of documentation on

>

> this topic?

>

> I'm trying to detect the transformation matrix between consecutive

>

> frames in a video with feature points detection (SIFT) and calculating

>

> the optical flow with the Lucas-Kanade method in pyramids.

>

> Therefore I used RANSAC to estimate the best omography transormation

>

> and added a layer of augmented reality with cvWarpPerspective.

>

> The main two techniques I found are the following:

>

> - FrameToReference: I get the feature points from the first frame

>

> (reference frame) and then compute the matches in the follwing frames

>

> with the first as reference. The problem here is that after a while

>

> the current frame could be very different from the reference one, and

>

> the matching of the correspondences too hard.

>

> - Frame to Frame: The matching of feature points is always calculated

>

> between consecutive frames. This could be simpler, because consecutive

>

> frames are very similar, but on the other hand, the homography is

>

> incremental and every little error in each step adds up; after a few

>

> seconds the omography is too wrong.

>

> Is there a way to mix up this two techniques to get better results?

>

> Thanks :)

>

>

>

>

>

>

>

>

>

>

>

> [Non-text portions of this message have been removed]

>

>


     

   
   
       
         
       
       








       


       
       


     

[Non-text portions of this message have been removed]

Reply | Threaded
Open this post in threaded view
|

Re: Image Registration and Augmented Reality

luca.dantona
In reply to this post by luca.dantona
Thanks to everyone for all the answers :)