Ask Your Question
0

object detection and tracking at known background

asked 2018-06-08 19:33:48 -0500

Sanglee gravatar image

Hi! I am trying to code to track an object. Here are some information.

  1. The background is complicating (many edges, objects) but mostly static (only slow changes such as sunlight). The object is mostly always moving and it moves much faster than the background does.
  2. The approx. size and shape of the object is known. It is a stick with a known size. So I can expect the min/max size of the object when it appears to the image.
  3. The stick (object) changes its location and often rotates a bit.
  4. The color of the stick is mostly known. But its color is similar to the background of it.

I have tried to train LBP, but it did not work well due to true negatives. (I even tried to train the background as for the negative images without the object). Thus, I moved to a very classic method. color filtering, edge detection, and etc. Probably it will first filter with a color range, remove blobs which are inside an object, find contours with a similar shape. Here my questions is...

  1. For removing blobs, I just have contours how. What should I do to fill small blobs? (_, contours, _) = cv2.findContours(frame, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
  2. I have a rectangle frame which overs the object which may be rotated. How could I calculate the ratio of width and height? I got the rectangle from ... rect = np.int32(cv2.boxPoints(cv2.minAreaRect(cnt)))
  3. The most important question is.... Do you think there is any good way/model to achieve my goal? Deep learning would not be possible considering the object's image is not easy to get.
edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted
0

answered 2018-06-11 21:25:37 -0500

Tetragramm gravatar image

Check out the Background Subtraction methods, like THIS tutorial.

That gives you a binary foreground/background separation. Then use morphological operations to clean up the binary mask a bit. Connected Components to get the blobs, then filter them by all of your features. Size, color, ect.

It should work pretty well.

edit flag offensive delete link more

Comments

Thank you very much for your reply! I tried it, but since the color of the object and the background are similar, the object is not easily realized. Will be there any way to maximize of detecting an object in such situation? or is there a parameter the enhances this performance of Background Subtraction models?

Sanglee gravatar imageSanglee ( 2018-06-13 10:32:13 -0500 )edit

Background detection is usually very good. Can you post a sample image or two to help us understand the problem?

Tetragramm gravatar imageTetragramm ( 2018-06-14 16:52:45 -0500 )edit
Login/Signup to Answer

Question Tools

1 follower

Stats

Asked: 2018-06-08 19:32:25 -0500

Seen: 54 times

Last updated: Jun 11