Why does the foreground image from background subtraction look transparent?

asked 2017-01-05 02:50:53 -0600

updated 2017-01-05 04:52:25 -0600

Hello!

I've been working on background subtraction techniques to detect and track humans in a scene. Because I don't want the background to change, and just subtract the frames from the first reference frame, so I've tried both the OpenCV MOG2(with learning parameter set to zero) and using the first frame and using the absdiff() to find the difference. But the foreground images I get from both techniques looks transparent as in, part of the background can be seen through the person as shown below:

Results with the absdiff() technique:

-Original Image:

image description

-Foreground Image:

image description

Results with the MOG2:

-Original Image:

image description

-Foreground Image:

image description

And this is the background(reference) image for both the methods:

image description

Does anyone know why this happens? I need to detect and track the people for which I find the blobs and because of the transparency the blobs are pretty much detected as two blobs which messes up everything else.

Thanks in advance!

edit retag flag offensive close merge delete

Comments

try without the black sweater ;)

(it can't see any difference between black chairs and black sweater)

berak ( 2017-01-05 03:10:09 -0600 )edit

Ah! Can't believe I hadn't thought of that! I feel silly now! Thanks anyways!

abhijith ( 2017-01-05 03:47:09 -0600 )edit

ofc, that's not the cure, just meant for a better understanding the situation.

actually, i'd think yours is quite a typical, real life case.

berak ( 2017-01-05 03:57:40 -0600 )edit

I know but this acts as another argument, a strong one, against not using background subtraction for human detection.

That is, of course, if there is no solution to overcome this problem right?

abhijith ( 2017-01-05 04:02:07 -0600 )edit

add a comment

Comments

This is what I'm working on. With tracking into play, I think I need some other heuristic especially if the shapes of my blobs change from one frame to another and the morphological operations don't always give a single blob.

Anyhoo, Thanks for your help!:)

abhijith ( 2017-01-05 05:01:38 -0600 )edit

maybe use those blobs only as a hint, which region to search with a more refined pass, e.g. take the boundingbox of it (findContours), and feed that roi to HOG person detection

berak ( 2017-01-05 05:08:35 -0600 )edit

I've been avoiding HOG because we have a whole bunch of processing on the detected person(age detection, gender detection, gesture detection) which are all, already computationally expensive and we need to run this code on a GPU. And also, HOG doesn't always detect people when only half their body is in frame

abhijith ( 2017-01-05 05:16:11 -0600 )edit

Your case is the exact reason why most academics working on pedestrian detection moved away from background subtraction and started using more advanced techniques like ICF, ACF, CNN's, ... I know they are computationally expensive, but most of them can already be heavily optimized to run real time!

StevenPuttemans ( 2017-01-05 06:04:33 -0600 )edit

I faced a similar situation i.e multiple blobs for same person, one above waist and another below. Use of morphological operations and filling with holes solved the problem.

ManuVISION ( 2017-01-10 08:39:55 -0600 )edit

add a comment

Why does the foreground image from background subtraction look transparent?

Comments

1 answer

Comments

Links

Question Tools

Stats

Related questions

Why does the foreground image from background subtraction look transparent? edit

Comments

1 answer

Comments

Links

Question Tools

Stats

Related questions

Why does the foreground image from background subtraction look transparent?