1

I need to develop an algorithm to remove a background from the foreground image. I would like to use chroma keying. This will be done on a live camera stream and I need to do it using OpenCV in C++.

The algorithm I thought would be something like:

  1. Capture frame from webcam into matrix
  2. Compute 3d rgb space graph
  3. The points around a certain value e.g. (10,255,10) for green background.
  4. Convert all points around green region to 0.
  5. And the rest to 1.
  6. Multiply that with the original image (convolution)
  7. Should get rid of background.

I would like some help on a method to remove a plain color background from foreground image without the color from the foreground image being removed.
Is there like a special type of crop function to remove background?

Bojangles
  • 99,427
  • 50
  • 170
  • 208
Seif Sharif
  • 103
  • 1
  • 2
  • 10

2 Answers2

0

I'm using OpenCV from Python. Suppose you followed a basic tutorial and have the frame from the camera and have a background image. A single line of code should do the trick with the numpy's where function.

outputImage = np.where(frame == (10,255,10), background, frame)

Now the outputImage contains the original frame where it wasn't green and your background image where it was green. As you said, you should use some tolerance.

b_m
  • 1,473
  • 2
  • 18
  • 29
  • so if i just search through the entire pixel and when a certain pixel is a certain value change it to 0, would that make it transparent. i don't know what type it is bmp or png as it is straight from the web cam. thanks for the hint. – Seif Sharif Feb 19 '13 at 21:47
  • no, if you change it to 0, it would be black. OpenCV cannot deal with transparency. You have to use masks for that. The example I given can change the green background of your scene to anything you like. – b_m Feb 19 '13 at 22:21
  • I want to capture an object and delete the background so after i can process the image. how would i implement your example into c++. I had a feeling i had to use mask, a mask would be a binary representation of the foreground image i want and i subtract that from the original image would that give me the image only. – Seif Sharif Feb 20 '13 at 17:26
  • so would this algorithm work store image in matrix; loop through matrix row by row incrementing the column number, check each element see if it matches the background color or in that range change it to zero, else change it to 1. this would produce the mask, then multiply the 2 and you should get the picture without the background – Seif Sharif Feb 21 '13 at 20:01
  • yes, it would. And It could be done without the nested loops with numpy with a single line of code much faster. – b_m Feb 22 '13 at 10:02
  • C++ doesn't recognized np.where function – Seif Sharif Mar 19 '13 at 19:16
0

I'm using OpenCV in C++, but I don 't fully understand what you want, @Seif (sorry for my poor English). Do you want to replace the green in the background with the colour of the foreground, and you're stuck at "Multiply that with the original image (convolution)"? Your algorithm is terribly true, but you must put the green to angle, circle... and implement them, there are many algorithms to do it (I'm doing it like that), otherwise, I think it's impossible except changing any green pixel to a similar pixel of your foreground.

SwiftsNamesake
  • 1,540
  • 2
  • 11
  • 25
Chien Proger
  • 1