Hi all,
I’m i need off some advise.
I’m studying the possibility to locate know objects (XY) in a picture.
You can call it bin picking, but then again (done with a lasergrid) it is not accurate enough, because of the low Z direction differences witch are only a few millimeters.
A picture off my task:
and the object to detect:
I’ve tried Tensorflow but i don’t know if that is the way to go for this project.
Any ideas on how to approach this matter?
Thank you
I’d recommend looking for commercial “machine vision” toolboxes/software packages. they regularly have to handle this task.
you’ll need higher resolution and better lighting so the edges of every object stand out better.
you want only topmost objects, I’m guessing, so be aware that some techniques will also detect occluded objects.
how do you want to pick these things? suction cup? magnet? grippers?
I’d recommend dumping them into a physical sorting mechanism. shake them up until they fall in line. using a robot and vision for such tasks is a waste of resources (also time for development).
Hi Crackwitz,
i’ve been looking at some commercial “machine vision” toolboxes. Those I’ve seen use a Laser grid, they work amazing on volume consuming pieces, but they struggle with the small high difference.
A higher resolution and better lighting, i agree. On my todo list.
Topmost objects, i wanna try with a timed flash, but i’m not sure it is going to work since high res camera use rolling shutters and these are to slow to ‘clip’ out the light.
the pick up will be done with suction cups, 1.5 folds so an angled object can be picked.
the shaking is my last option, so not yet
you gotta realize, I don’t mean “more expensive flash”.
I mean you should try to put the light elsewhere. like off to the side, and all around the box, so the surfaces of your parts don’t reflect (as they do now), but all the edges bounce some light into the camera.
that’s also what I mean by “commercial”. these companies take your money so they can tell you these fundamental things (and how to use their software so it recognizes your parts in a pile).
a decent laser scanner might still be required.
or you go the mechanical separation route, as mentioned above. that’s robust. that’s what people do when they want this to work reliably and they don’t have the patience or the gambling addiction to solve it any other way.
like this, but appropriately different. find a mechanical engineer and tell him to work the problem. the solution will be a lot simpler and faster than robotic pick and place…
