background subtraction

Re: background subtraction

"wingraider"  wrote in message 
news:c0f71bf8-dadd-4784-9be6-2b9577abc83a@i24g2000prf.googlegroups.com...
On Dec 9, 11:39 am, ImageAnalyst  wrote:
> On Dec 8, 7:09 pm, wingraider  wrote:
>
>
>
> > On Dec 9, 1:19 am, Martin Leese 
> > wrote:
>
> > > wingraider wrote:
> > > > can anyone give me a good robustbackgroundsubtractionwith easy code
> > > > implementation?
>
> > > for each pixel
> > > {
> > > foreground[pixel] = input[pixel] -background[pixel];
>
> > > }
>
> > > --
> > > Regards,
> > > Martin Leese
> > > E-mail: ple...@see.Web.for.e-mail.INVALID
> > > Web:http://members.tripod.com/martin_leese/
>
> > uhmm..thanks for the reply..i have tried that one,but it doesn't work
> > for me,I'm referring to a foreground cut off type rather than simple
> > differencing.
>
> > ex.
>
> > foreground:
>
> > yyyyyyyyyXXXXXXyyyyyyyyyy
>
> >background:
>
> > yyyyyyyyyyyyyyyyyyyyyyyyy
>
> > result
>
> > ----------------XXXXXX------------------
>
> -------------------------------------
> After inspecting your algorithm, I've come up with:
> for all x,y:
> if foreground(x,y) !=background(x,y)
> result(x,y) = foreground(x,y)
> else
> result(x,y) = undefined or null
> endif
> end for
>
> If you're trying to do something like detect a person walking in front
> of a static scene (such as a seated crowd and other static clutter)
> then this kind of thing (backgroundsubtraction) doesn't work because
> the "foreground" pixels are NOT ADDED to thebackgroundbut instead
> REPLACE it. So you'd end up with a very broken foreground with holes
> and corrupt outline.
>
> Regards,
> ImageAnalyst

>yes, I'm trying to detect  an object from a static background.
>the static  background is captured without any objects. then i'll put
>an object in the scene then capture the scene with an object.
>
>i want to extract the object from the recently captured scene with the
>use the background that was captured earlier but simple subtraction
>won't work for me. i get a very corrupted image out of it.

If these images are captured then you're going to get subtle differences in 
the backgrounds. This would be due to the tiniest camera shake, and most 
importantly the type of compression the image has undergone to get to the 
computer. Also, if a limit is placed on the amount of data because of 
streaming video, then a single change in a part of the image could effect 
the whole image.
You could check for pixels that are NEARLY the same within some threshold, 
or try filtering the images before the testing to removing any high 
frequency noise.

Re: background subtraction

Hello,

Once I was writing some motion detection algorithms and one of them
was based on similar technique =96 subtracting background from current
video frame to locate moving object. You may check some ideas on Code
Project:
http://www.codeproject.com/KB/audio-video/Motion_Detection.aspx

To do what you need (get an object) you may need differencing,
thresholding and masking:
1) Get difference image:
for each pixel
{
     difference[pixel] =3D input[pixel] - background[pixel];
}

2) Threshold difference
for each pixel
{
     difference[pixel] =3D (difference[pixel] > threshold ) ? 1 : 0;
}

3) Masking =96 remove background from input image.
for each pixel
{
    If (difference[pixel] =3D=3D 0 )
        input[pixel] =3D 0;
}

Since you may get some noise from your camera, I would suggest using
some noise suppression techniques between 2nd and 3rd steps. For
example math morphology could work (erosion, opening).

With best regards,
Andrew

http://www.aforgenet.com




>
> yes, I'm trying to detect =A0an object from a static background.
> the static =A0background is captured without any objects. then i'll put
> an object in the scene then capture the scene with an object.
>
> i want to extract the object from the recently captured scene with the
> use the background that was captured earlier but simple subtraction
> won't work for me. i get a very corrupted image out of it.- Hide quoted t=
ext -
>
> - Show quoted text -

Re: background subtraction

On Dec 12, 3:35=A0am, Andrew Kirillov  wrote:
> Hello,
>
> Once I was writing some motion detection algorithms and one of them
> was based on similar technique =96 subtracting background from current
> video frame to locate moving object. You may check some ideas on Code
> Project:http://www.codeproject.com/KB/audio-video/Motion_Detection.aspx
>
> To do what you need (get an object) you may need differencing,
> thresholding and masking:
> 1) Get difference image:
> for each pixel
> {
> =A0 =A0 =A0difference[pixel] =3D input[pixel] - background[pixel];
>
> }
>
> 2) Threshold difference
> for each pixel
> {
> =A0 =A0 =A0difference[pixel] =3D (difference[pixel] > threshold ) ? 1 : 0=
;
>
> }
>
> 3) Masking =96 remove background from input image.
> for each pixel
> {
> =A0 =A0 If (difference[pixel] =3D=3D 0 )
> =A0 =A0 =A0 =A0 input[pixel] =3D 0;
>
> }
>
> Since you may get some noise from your camera, I would suggest using
> some noise suppression techniques between 2nd and 3rd steps. For
> example math morphology could work (erosion, opening).
>
> With best regards,
> Andrew
>
> http://www.aforgenet.com
>
>
>
>
>
> > yes, I'm trying to detect =A0an object from a static background.
> > the static =A0background is captured without any objects. then i'll put
> > an object in the scene then capture the scene with an object.
>
> > i want to extract the object from the recently captured scene with the
> > use the background that was captured earlier but simple subtraction
> > won't work for me. i get a very corrupted image out of it.- Hide quoted=
 text -
>
> > - Show quoted text -- Hide quoted text -
>
> - Show quoted text -

---------------------------------------------------------------------------=
-
Andrew:
And what did you do when the foreground object was darker than your
background, or the same intensity as your background?  You'd have
holes in your thresholded object.  Did you simply do a hole fill?
Does that work OK?  What about if those regions are at the edge of the
object so that a hole fill won't work?  What did you do then?
Regards,
ImageAnalyst

Re: background subtraction

Sorry, I did not make my difference idea clear. I take absolute
difference, which is |input - background|. So even if object is
darker, it is not issue. The issue appears only when object's
intensity is close to background. Then the result will be as you've
described.

With best regards,
Andrew

http://www.aforgenet.com

>
> -------------------------------------------------------------------------=
--=AD-
> Andrew:
> And what did you do when the foreground object was darker than your
> background, or the same intensity as your background? =A0You'd have
> holes in your thresholded object. =A0Did you simply do a hole fill?
> Does that work OK? =A0What about if those regions are at the edge of the
> object so that a hole fill won't work? =A0What did you do then?
> Regards,
> ImageAnalyst- Hide quoted text -
>
> - Show quoted text -

Re: background subtraction - with shadows

Andrew, thanks for the good concise algorithm.
I am doing about the same thing, with decent results.


I am building a scorekeeping system for car races, using video input
to detect the cars as they cross the finish line, recognize each car
by car number, color, and any other features I can find, and keep
track of times and lap counts for each car.  Currently, all major car
races (NASCAR, Indy car, Formula 1) require the cars to carry radio
transponders to identify the cars to a scoring computer, which is
expensive and trouble-prone.  I'm trying to use video recognition to
replace the transponders.
I am using the PIL (Python Image Library) for my project; it's easy to
use and free, but doesn't have as many features as other packages such
as 'blob detection', and Python is significantly slower than other
software, so I may have a challenge working in real-time.


The first few steps of my process are:
1) capture an empty 'background' image (and refresh it on occasion)
2) for each subsequent frame, get a 'difference' image
3) use a few filters to remove 'noise', adjust to changing light
conditions, etc.
4) if the pixel count is less than X, there's 'nothing more to see
here'; wait for the next frame; otherwise ...
5) generate a 'mask' from the 'difference' image
6) 'solidify' the mask: fill in holes and crevasses
7) use the mask to cut-and-paste the new object (one or more cars, in
my case) from the scene.

This works pretty well ... but I am not doing a very good job of
identifying, and masking off, shadows, which of course give misleading
information about the car's size, shape, and colors.  I have some ugly
code to compare the edges of the mask to the background: if a pixel is
X percent darker than the background, it *might* be a shadow.

Has anyone else done a decent job of separating features from shadows?

Gary Davis
earth to gary at earth link period n e t


On Fri, 12 Dec 2008 00:35:15 -0800 (PST), Andrew Kirillov
 wrote:

>Hello,
>
>Once I was writing some motion detection algorithms and one of them
>was based on similar technique – subtracting background from current
>video frame to locate moving object. You may check some ideas on Code
>Project:
>http://www.codeproject.com/KB/audio-video/Motion_Detection.aspx

[snip]

Re: background subtraction - with shadows

On Dec 13, 1:12=A0pm, Pioneer  wrote:
> Andrew, thanks for the good concise algorithm.
> I am doing about the same thing, with decent results.
>
> 
> I am building a scorekeeping system for car races, using video input
> to detect the cars as they cross the finish line, recognize each car
> by car number, color, and any other features I can find, and keep
> track of times and lap counts for each car. =A0Currently, all major car
> races (NASCAR, Indy car, Formula 1) require the cars to carry radio
> transponders to identify the cars to a scoring computer, which is
> expensive and trouble-prone. =A0I'm trying to use video recognition to
> replace the transponders.
> I am using the PIL (Python Image Library) for my project; it's easy to
> use and free, but doesn't have as many features as other packages such
> as 'blob detection', and Python is significantly slower than other
> software, so I may have a challenge working in real-time.
> 
>
> The first few steps of my process are:
> 1) capture an empty 'background' image (and refresh it on occasion)
> 2) for each subsequent frame, get a 'difference' image
> 3) use a few filters to remove 'noise', adjust to changing light
> conditions, etc.
> 4) if the pixel count is less than X, there's 'nothing more to see
> here'; wait for the next frame; otherwise ...
> 5) generate a 'mask' from the 'difference' image
> 6) 'solidify' the mask: fill in holes and crevasses
> 7) use the mask to cut-and-paste the new object (one or more cars, in
> my case) from the scene.
>
> This works pretty well ... but I am not doing a very good job of
> identifying, and masking off, shadows, which of course give misleading
> information about the car's size, shape, and colors. =A0I have some ugly
> code to compare the edges of the mask to the background: if a pixel is
> X percent darker than the background, it *might* be a shadow.
>
> Has anyone else done a decent job of separating features from shadows?
>
> Gary Davis
> earth to gary at earth link period n e t
>
> On Fri, 12 Dec 2008 00:35:15 -0800 (PST), Andrew Kirillov
>
>  wrote:
> >Hello,
>
> >Once I was writing some motion detection algorithms and one of them
> >was based on similar technique =96 subtracting background from current
> >video frame to locate moving object. You may check some ideas on Code
> >Project:
> >http://www.codeproject.com/KB/audio-video/Motion_Detection.aspx
>
> [snip]

--------------------------------------------
Pioneer:
Just 3 days ago I answered something like this in the MATLAB
newsgroup:
http://tinyurl.com/5uvtvh
I think that method could work but you need to look at surrounding
pixels and see if they are all down by the same percentage.  Because
otherwise, ANY dark pixel could be a percentage of the background (for
example, it a black car came into view).  That's why you need to check
a bunch of surrounding pixels to see if they are all about the same
percentage down.  But otherwise, with a few tweaks and things, it
could work.
Regards,
ImageAnalyst

Re: background subtraction - with shadows

Pioneer wrote:

> Has anyone else done a decent job of separating features from shadows?

Have you considered using a calibrated, stereo vision system? Shadows
always stay on the ground, but most pixels on the car will be above
the ground. 3D might help your classification, too, for example you
could use depth discontinuities to tell whether your blob was a single
car or one car occluding another one.

Of course, you'd have to solve the correspondence problem, but I'd have
thought that the finish line around a race track would have plenty
of good features to use as markers.

Cheers

mark-r

Re: background subtraction

On Dec 12, 4:35=A0pm, Andrew Kirillov  wrote:
> Hello,
>
> Once I was writing some motion detection algorithms and one of them
> was based on similar technique =96 subtractingbackgroundfrom current
> video frame to locate moving object. You may check some ideas on Code
> Project:http://www.codeproject.com/KB/audio-video/Motion_Detection.aspx
>
> To do what you need (get an object) you may need differencing,
> thresholding and masking:
> 1) Get difference image:
> for each pixel
> {
> =A0 =A0 =A0difference[pixel] =3D input[pixel] -background[pixel];
>
> }
>
> 2) Threshold difference
> for each pixel
> {
> =A0 =A0 =A0difference[pixel] =3D (difference[pixel] > threshold ) ? 1 : 0=
;
>
> }
>
> 3) Masking =96 removebackgroundfrom input image.
> for each pixel
> {
> =A0 =A0 If (difference[pixel] =3D=3D 0 )
> =A0 =A0 =A0 =A0 input[pixel] =3D 0;
>
> }
>
> Since you may get some noise from your camera, I would suggest using
> some noise suppression techniques between 2nd and 3rd steps. For
> example math morphology could work (erosion, opening).
>
> With best regards,
> Andrew
>
> http://www.aforgenet.com
>
>
>
> > yes, I'm trying to detect =A0an object from a staticbackground.
> > the static =A0backgroundis captured without any objects. then i'll put
> > an object in the scene then capture the scene with an object.
>
> > i want to extract the object from the recently captured scene with the
> > use thebackgroundthat was captured earlier but simplesubtraction
> > won't work for me. i get a very corrupted image out of it.- Hide quoted=
 text -
>
> > - Show quoted text -

thanks andrew for that logic flow,i'm now having trouble on threshold
selection.
is there any way to automatically select thresold value without going
to grayscaling the image and finding the minima and maxima of the
histogram?