|
Science Forum Index » Image Processing Forum » Pyramidial Lucas Kanade Optical Flow Estimation
Page 1 of 1
|
| Author |
Message |
| Guest |
 Posted: Thu Dec 14, 2006 6:21 am |
|
|
|
|
Hi,
I use to segment moving objects from the background by estimating the
optical flow with the pyramidial lucas kanade algorithm. But there is a
problem with the big samplewindow i have to use to make it robust.
Because of that window by a smooth background the calculatet flow field
is bigger than the moving object.
1. Are there some other algorithm of flow estimation that didn't have
this problem?
2. or are there variations of the pLK algorithm, who eliminate these
errors
It is important that the algorithm can handle displacements of more the
1 pixels.
I would be glad of some other experiences with the pyramidial lucas
kanda or the segmentation with optical flow by big displacements und
high accuracy |
|
|
| Back to top |
|
| Abhishek |
| Posted: Fri Dec 15, 2006 12:47 pm |
|
|
|
Guest
|
Hi,
What do u mean by the sample window in Lukas Kanade Optical
Flow?
Are u referring to the Aperture problem?
With Regards,
Abhishek S
Tobias.Senst@student.uni-magdeburg.de wrote:
Quote: Hi,
I use to segment moving objects from the background by estimating the
optical flow with the pyramidial lucas kanade algorithm. But there is a
problem with the big samplewindow i have to use to make it robust.
Because of that window by a smooth background the calculatet flow field
is bigger than the moving object.
1. Are there some other algorithm of flow estimation that didn't have
this problem?
2. or are there variations of the pLK algorithm, who eliminate these
errors
It is important that the algorithm can handle displacements of more the
1 pixels.
I would be glad of some other experiences with the pyramidial lucas
kanda or the segmentation with optical flow by big displacements und
high accuracy |
|
|
| Back to top |
|
| J.Senst |
| Posted: Sat Dec 16, 2006 9:37 am |
|
|
|
Guest
|
With the sample window I mean the Region where one flow-vector will be
calculated. By the Pyramidial Lucas Kanade Algorithm you have two
parameters the level of the pyramid structure and the size of the
window for that a vector will be calculated.
And if in one "big" window are 2 moving objects for example one car
moving from the left to the right and an other from the right to the
left, you got an error of the calculated flow-vector.
So it is not a problem of aperture more an problem like undersampling .
With Regards,
J.Senst |
|
|
| Back to top |
|
| Abhishek |
| Posted: Sat Dec 16, 2006 12:14 pm |
|
|
|
Guest
|
Hi,
Yes you are talking of the aperture problem, and to solve that
problem, they consider the flow to satisfy more than one pixel,
actually all the pixels in the window. Now, why dont u use a small
window size of may be 5X5 or 3X3? What tool are u using to calculate
the flow?
As u move below the pyramid, gradually increase the window size
( I would advice u to double it as the rows and cols also double).
Does it clear your point or are you still in confusion?
With Regards,
Abhishek S
J.Senst wrote:
Quote: With the sample window I mean the Region where one flow-vector will be
calculated. By the Pyramidial Lucas Kanade Algorithm you have two
parameters the level of the pyramid structure and the size of the
window for that a vector will be calculated.
And if in one "big" window are 2 moving objects for example one car
moving from the left to the right and an other from the right to the
left, you got an error of the calculated flow-vector.
So it is not a problem of aperture more an problem like undersampling .
With Regards,
J.Senst |
|
|
| Back to top |
|
| J.Senst |
| Posted: Tue Dec 26, 2006 6:18 am |
|
|
|
Guest
|
Hi,
Thanks!! I used the implementation of OpenCV. The problem I have with
these small window sizes is that my source images have mostly at the
important place not enough texture, so that the estimation result is
incorrect and i need a bigger window like the size of 15x15
With Regards,
J.Senst |
|
|
| Back to top |
|
| |
