Machine Vision

cancel
Showing results for 
Search instead for 
Did you mean: 

Counting TMT

Hello everyone.

I am looking to count the no.of TMT bars in the bundle. The problem I am facing are -

1) At times few components from the background also get counted, like the bright spot in the window.

2) Many of the TMT's are hidden behind the other bars, and so a very smal curvature / area is exposed whihc makes it tough to identify.

3) When I try thresholding, at many places two TMT bars which are very closely placed to each other get their faces ingled as a result of which the area of two TMT may be considered as one.

 

I was using vision builder (VBAI) for this porpose. Have a look at the below attached images and share your thoughts.

 

 

Also I have a problem when I change the image, the threshold values of the first image oes not give a satisfactory result.

0 Kudos
Message 1 of 7
(5,708 Views)

I don't think I would take this project on.  It looks close to impossible to me.  I would think weighing the bars would produce much better results, or counting accurately as they are produced.

 

If I had to do it, I would probably try to back the camera up and use a larger focal length lens.  This would reduce the perspective issues (overlapping ends, etc.).

 

Getting a good shiny spot for each end is critical, also.  A large light would hit all the ends so that they reflect into the camera.  It would probably need to be about twice the size of the pile of bars.

 

If you had bright spots with a darker background, you could use dynamic thresholding to look for spots of a certain size.

 

Bruce

Bruce Ammons
Ammons Engineering
0 Kudos
Message 2 of 7
(5,689 Views)

I agree with Bruce, this is a difficult application, especially if you want a spot on / exact count of the bars.

If you are fine with a more fuzzy count, you might want to try to implement the algorithm described in this paper:

http://www.robots.ox.ac.uk/~vgg/research/counting/

 

I've seen good results using this method on images that are otherwise difficult to process using regular segmentation methods. I hope this can be helpful in your specific application.

 

Best regards,

 

Christophe

 

0 Kudos
Message 3 of 7
(5,672 Views)

How are you imaging this bundle.  Lighting and imaging technique will greatly effect your success.  If you have poor control of this you will have some difficulty using a simple threshold and blob counting. Your best bet is to control the imaging so that you can get a uniform reflection on the surfaces normal to the camera, the surfaces you want to count.  I have done many counting applications, one other trick is to use more than one image to get more views to work from.  And yes backing up the camera should help with blocked views due to perspective.

 

 

 

 

Paul Falkenstein
Coleman Technologies Inc.
CLA, CPI, AIA-Vision
Labview 4.0- 2013, RT, Vision, FPGA
0 Kudos
Message 4 of 7
(5,650 Views)

But how to implement this method? I have tried the learning in 'Clascify Objects ' in the Vision Builder Software.. but it is very strict i.e. if I have made it learn 5 obects , it'll only detect those objects which are exactly equal to those 5 objects.

 

How to implement a custom algorithm?

0 Kudos
Message 5 of 7
(5,623 Views)

What do you mean when you say a "more fuzzy count" ? does that mean more unreliable??

0 Kudos
Message 6 of 7
(5,579 Views)

Hi , Would like to know, did you solve the problem? 

 

 

0 Kudos
Message 7 of 7
(2,646 Views)