• Advertisement

Archived

This topic is now archived and is closed to further replies.

Objects in images

This topic is 5773 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

hi, i''m lookin for a short way to find and recognize objects on a picture. I just want to figure out what''s a human and what''s a desk. the algorithm hasn''t got to be perfect, but small and fast. is there anybody having an idea? cYa DjR

Share this post


Link to post
Share on other sites
Advertisement
Hehe.. smal and simple... :-)
I think there are one hundrer ways to try to do this... the best one would be to combine maybe 20 of them... :-)
This is a aproach i always wanted to test:

First step is to extract information out of your image!

Make a grayscale version of the image, make one with only the colors, not the shades.. (RGB 128,64,0 >> RGB 256,128,0)
then you need a frame memory with maybe 4-5 frames... with these, you test the differense your frame have with all the older frames to find movement...

NOW you have one grayscale image, one color image and one movement image!

Then you create 3 grayscaled contrast images out off these three images with a contrast filter! (simple matrix calc)

You now got good information on color , shade and movement!
You should make a filter witch sets a treshold on all 3 images, so that the result is 3 black and white images! (edge/noedge)

The whole point of image rekognition is to simplefy the information from 150kbs to :"TABLE & THREE" :-)

The next step, witch is the hard one is to add lines to your image... find places where the black and white pictures have black pixels forming a line! Try to make a algorithm so that maybe only the 100 strongest lines are saved... you should store a number with each line to remember the power of the line... thrash ewery line shorter then 3-10(???) pixels...

Then you try to compear the lines from the three b/w images, and store them in different places, like:
movement , shade , color , movement&shade , movement&color shade&color , movement&shade&color

NOW :-) You got to learn your computer what a table is, and what a person is :-) HEHE.. extreamely hard!.... You can either hard code it, or you can realy make it learn(super hard)!
Then we have a desc! It should have some longe strait lines along the edges, and it should not move! :-)
A face is a place with aloth of shade contrast, but little color contrast, and aloth of movement! ... :-)

The last two examples are realy simple, but by looking at your data, you should find patterns witch you could test, and then hard code, so that it recognises different things...
You can later make it find faces, and then find points to messure in the face and by time make it recognise differences in faces... :-) long shot i think...
You have to use the statistic aproach... like something move aloth , something have aloth of contrast , something dont move,
and so on...

It should also remember what it found for the last 15-20 frames, so that it it did not find a person in one frame, it maybe wait 2-5 seconds before eliminating the person from the what i see list :-)

Image recognition is hard.. and the only way is to simplefy and make abstract simplifications of what you see... ! The recognition part must either be hard coded (little flexible) or learned (slow)...

good luck :-)



Share this post


Link to post
Share on other sites
Its difficult to say without seeing the images.

If you''re talking about a general solution to pattern classification then the solution is not going to be a ''small and fast''!

Problems such as these are very difficult to solve using current machine vision techniques.

As I said though, it depends on the images and their composition.

Share this post


Link to post
Share on other sites
yes, we really need to see some samples before giving advice. This problem is pretty complex.

1. Do the images only have either a table or a human?

2. Are all the images lit the same?

3. Is the background uniform?

4. Do the images move?






Stimulate

Share this post


Link to post
Share on other sites
hmm..
the images aren't really moving - but the camera does.

I just want to figure out, what kind of computer power has to be used for some task like that. I want to know whether a small cpu can take those tasks.
so it's important to recognize obstacles as well as human beings or animals. I don't want to know who the human is (not yet) but that it is a human...
a picture of 64*64 Pixels might be enough. ?
(LOL does anybody know where i can get a ccd-chip with this resolution?)



[edited by - DocJunioR on June 28, 2002 12:45:27 PM]

Share this post


Link to post
Share on other sites
Guest Anonymous Poster
Object Recognition is an active area of research in the machine vision community, and there are no easy or general solutions as Carrot has pointed out. In addition, many systems rely on a vast amount of training data, which is then compressed into some sort of representation that can be stored (such as principle components) and used later to get a best match in the object recognition system. The problem becomes much easier if your problem domain is limited, but small and fast is still a luxury that the computer vision community is searching for...

Share this post


Link to post
Share on other sites
Just to add to AP''s response. If the camera is moving then the image you want to analyze is also moving. This increases the complexity of the problem enormously because you have to deal with areas of interest . A good man to ask about this type of problem is motters. You’ll find him at the generation5.org forum. He is developing a vision system for his robot.

Why do you want to do this anyway? Is it for a game? Give us some more details as there may be a much simpler way of approaching the problem.





Stimulate

Share this post


Link to post
Share on other sites
I''m just thinking about to build a robot. I just want to play around with some possibilities.
But small robots can only have small cpu''s so it''s gotta be a small vision system

Share this post


Link to post
Share on other sites
quote:
Original post by DocJunioR
I'm just thinking about to build a robot. I just want to play around with some possibilities.
But small robots can only have small cpu's so it's gotta be a small vision system



In which case it's going to be a simple/slow vision system, without much ability to recognise/classify objects.

Realistically, this is not the appropriate forum for this discussion (about robots), since GameDev is about game development. Of course, this sort of problem could be related to artificial agents playing in a simulated world and having to recognise objects... but its far easier just to tell the agent what the object is!

Bottom line is, that what you are trying to do is still very difficult, even for the top researchers in the field who have big research budgets, lots of grad students to help out and lots of computing power at their disposal. But that's not to say that you shouldn't try!

If you want to try this and you want/need to keep it small, I suggest that you actually build a subsumption architecture based vision system. Several sensors recording different information and each guiding the robots actions independently based on what they see/hear/detect. The trick is to build a system of independent controllers that actually work together to guide your robots actions!

Cheers,

Timkin


[edited by - Timkin on June 29, 2002 8:52:41 PM]

Share this post


Link to post
Share on other sites
quote:
Original post by DocJunioR
I''m just thinking about to build a robot. I just want to play around with some possibilities.
But small robots can only have small cpu''s so it''s gotta be a small vision system




Then I really would suggest you go talk to motters. He has much experience building robots and I''m sure he''ll help you get started. He also has a website with useful information and downloads of his image processing systems.

Meet Rodney, his robot..

http://www.fuzzgun.btinternet.co.uk/rodney/rodney.htm






Stimulate

Share this post


Link to post
Share on other sites

  • Advertisement