Ask Your Question
1

Counting Pallet Object on Warehouse

asked 2017-08-01 22:30:03 -0500

NPE gravatar image

updated 2017-08-30 23:48:48 -0500

Hi,

Recently working on project that want to count pallet on warehouse, the goals is to know how many pallet on warehouse.

Please refer my picture bellow :

image description

To simplified this problem, we create a sample area. Sample area is area that we need to know : is in the area that are pallet inside?

So, our problem was simplified to picture bellow :

image description

While Yellow indicate sampling area.

Thus, we have tried using canny algorithm or hough algorithm. No ones give me a best result. Sometimes give us wrong result like this bellow :

image description

While red indicates this is empty space,
While green indicates this area occupied by box pallet.

The problem our algorithm not working well in every scenario (time,pallet position, lighting problem, etc).

image description

As you can see (Row 6 Col 2), if the sampling area on the lighting spot then algorithm marking it's area occupied. this is totally wrong.

Lighting problem image example :
lighting problem

So, is there any idea or library on open CV that we can use, adopt, and integrate for our case? Is there any way to avoid the lighting problem?

Thanks in advance.

edit retag flag offensive close merge delete

Comments

could you post this image with original size

image

sturkmen gravatar imagesturkmen ( 2017-08-01 23:03:14 -0500 )edit

Off course, here you are : https://goo.gl/PqvLAT

For information, this image generated by stitching process.

NPE gravatar imageNPE ( 2017-08-01 23:18:58 -0500 )edit

nice project! if possible, also i want to see some of images which used to generate this image ( they have more detail ). you can do counting process before stitching

sturkmen gravatar imagesturkmen ( 2017-08-02 06:22:14 -0500 )edit

Alternative method, maybe too naive, use a background subtraction like method with as reference an image without the pallets? The constraints are a fixed camera, the same reference background and no / few other perturbations.

Eduardo gravatar imageEduardo ( 2017-08-02 07:31:35 -0500 )edit

sure, here you are : https://goo.gl/jmeXac

NPE gravatar imageNPE ( 2017-08-02 22:32:14 -0500 )edit

Is there any detail paper or project that can be reference for us?

NPE gravatar imageNPE ( 2017-08-02 22:40:37 -0500 )edit

i want to think on how to help you. could you post at least two ( ideally four ) images used to generate big picture

sturkmen gravatar imagesturkmen ( 2017-08-03 09:31:38 -0500 )edit

Of course, here you are : https://goo.gl/5pMh6U recently, i added 4 images. Did you know how to reduce lighting noise problem?

NPE gravatar imageNPE ( 2017-08-03 20:59:24 -0500 )edit

How i can avoid lighting problem in our scenario? is there any idea?

NPE gravatar imageNPE ( 2017-08-30 23:44:22 -0500 )edit

@NPE i am really wondering why you are simply ignoring the fact that I told you to move on to machine learning ... its the only way to robustly cope with changing lighting conditions...

StevenPuttemans gravatar imageStevenPuttemans ( 2017-08-31 03:29:01 -0500 )edit

1 answer

Sort by ยป oldest newest most voted
1

answered 2017-08-02 06:51:09 -0500

Instead of using canny or hough algorithm, generate a global descriptor per patch and use any machine learning mechanism, like a support vector machine, to separate pallet from non pallet patches. When choosing a global feature descriptor, look at LBP histograms or so, but select a feature descriptor that is known for a good texture classification!

edit flag offensive delete link more

Comments

Can you provide more detail? some sample project or paper that nice to read.

NPE gravatar imageNPE ( 2017-08-02 22:30:45 -0500 )edit

Ow well I do not know of a paper that does something similar, but about texture classification using OpenCV and LBP try this: http://www.pyimagesearch.com/2015/12/... I know it is in python, but it basically tells you what I suggested and how to apply it. Retrieving the patches yourself should not be the issue.

StevenPuttemans gravatar imageStevenPuttemans ( 2017-08-03 02:41:37 -0500 )edit

Obviously, problem comes where on lighting spot. Did you know how to reduce lighting effect noise on image?

NPE gravatar imageNPE ( 2017-08-03 20:55:31 -0500 )edit

How about when you also add these lighting spots in your training data? You need to cover as much conditions as possible. Histogram equalization on the patches might help also!

StevenPuttemans gravatar imageStevenPuttemans ( 2017-08-04 06:03:26 -0500 )edit

Obviously we don't use a deep learning algorithm. So we don't have any training data set. We just using canny algorithm to find the edges. And do simple calculation. If edges in one area more than X number of threshold (eg. 15), than this area contains a pallete. Is there any idea to improve accuracy our algorithm?

NPE gravatar imageNPE ( 2017-08-06 21:49:14 -0500 )edit

@NPE, that is what I am saying, you do not need to use deep learning at once, but you could simply collect 100 patches of each class. Your algorithm is manual thresholding and that is proven to almost always fail or break down.

StevenPuttemans gravatar imageStevenPuttemans ( 2017-08-07 09:07:07 -0500 )edit
Login/Signup to Answer

Question Tools

1 follower

Stats

Asked: 2017-08-01 22:30:03 -0500

Seen: 784 times

Last updated: Aug 30 '17