BorlandTalk.com

Jorrit Jongma · Guest

Hi group,

I need to do some video recognition, I have never done this before so I am
wondering where and how to start.

I thought of handling the video's like this:
- Break down the frames from the original video into a collection of objects
(detectable forms by using edge/corner recognition)
- Store this information

When new video plays:
- Break down the frames of the new video, see if they match
- Compare to the original stored data, with a bit of fuzzyness since the
original and new video can be of somewhat different image quality

This was all I could come up with to do it. Is this the right way to go?
Ofcourse then we have the matter of speed. The original videos will be
around 200, so that isn't to many. Will it be possible to reach realtime
comparing speed with the new playing video? Breaking down and comparing do
not necessarily have to be done on the same PC. They may both be done on
dedicated multiple proc machines with quite some memory.

Pointers, information, tales of experience, all is welcome!

TiA

Roy Magne Klever · Guest

Hi Jorrit,

If you go to www.delphipages.com and do a search for TMED you will get a
freeware component for doing Motion and Edge detection. It is very good.

Kind Regards
Roy Magne Klever

Nils Haeck · Guest

What do you want to recognise in the video?

Or do you mean: video comparison?

There are many different approaches and concepts. Just to name one example:
facial recognition is usually done by transforming pictures into an
eigenvalue problem and finding the eigenvalues that are characteristic for
faces. This is quite a complicated process already.

Nils
www.simdesign.nl

"Jorrit Jongma >" <

Jorrit Jongma · Guest

Video comparison might be a better description.

I have one video stream, embedded in this stream is another video stream I
already have on file. I want to know when this embedded stream plays.

Nils Haeck · Guest

Aha, that sounds encouraging.

So, suppose that you can perfectly estimate if two frames are identical.

Lets consider your one video stream to be built of frames:
Stream 1: ABCDEFGHIJKLMNOPQR

Now, you also have another stream somewhere, that has

Stream 2: HIJKLMN

and

Stream 3: FZWXY

You can start with frame H of stream 2, then compare it to the frames in
stream 1. Obviously, the first ABCDEFG frames are no matches. Then, you find
that H matches H.

The next step is to find if following frames also match. Yes they do, we
find I=I, J=J, etc until N=N.

When we want to find stream 3, we will find a match with F = F. But, in this
case, following frames do not match, after all Z <> G, W <> H etcetera.

In order to use above, you must be certain that the videos have the exact
same frame rate and timing.

You can make the comparison quite relaxed, after all, you are comparing
quite a few frames probably, and so you can afford having a few false
positives.

What I would do is probably create a thumbnail image, something like 4x4
pixels, of each frame. Compare these 16 pixels with 16 pixels of the other
frame. If their differences (absolute values, added together) are below a
certain treshold, you consider the frames equal.

This step is a little bit like the first step in my image comparison
function in ABC-View Manager (a shareware viewer that can find similar
images). Perhaps nice to try for you:
http://www.abc-view.com/abcview.html
http://www.abc-view.com/articles/article3.html

Hope that helps,

Nils Haeck
www.simdesign.nl

"Jorrit Jongma >" <

Monty · Guest

Nils, i have checked your image comparison function and it's great!

Are you planning to sell this feature as a separate component?

"Nils Haeck" <n.haeckno (AT) spamchello (DOT) nl> escribió en el mensaje
news:40749502$1 (AT) newsgroups (DOT) borland.com...

Nils Haeck · Guest

Jorrit Jongma · Guest

Nils Haeck · Guest

I already gave you some advice.. Divide a frame into 16 blocks (4x4). Each
of these blocks has a colour, but let's do this for grayscale as an example.

There are block values X[i], where i = 0..15 (16 values) and Y[i] for
another frame, always a value from 0 to 255.

Now, your error (per block) = abs(X[i] - Y[i])

The total error:

Err = Sigma(abs(X[i] - Y[i]), i = 0..15

or with code

Err := 0;
for i := 0 to 15 do
Err := Err + abs(X[i] - Y[i]);

Of course, the images are not exactly the same, but images that are close
will have a small total error. So you can say

IsIdentical := Error < Limit;

Experiment with a few limit settings.

Search back in this newsgroup to find fast downscale algorithms you can use
to scale (any) image down to 4x4 size.

Kind regards,

Nils Haeck
www.simdesign.nl

"Jorrit Jongma >" <

shahram · Guest

Hi,
Does your project depend on video file format or not (your video file is
MPEGII, MPEG4 or non-standard)? I think if you use the data and information
that exists in the file, you will do half of your project. After that you
compare two digitized frame.

"Jorrit Jongma >" <

Jorrit Jongma · Guest

Dominique Louis · Guest

Jorrit Jongma < wrote: