Wanted to buy: Moderate pick and place line.

[Kjelt]

Just wondering: is there a universal programming language to control P&P machines, such as the G-codes for CNC machines, or is it every manufacturer has its own standard?

[mikeselectricstuff]

   Its amazing that a raspberry PI can do the Vision.  It does it, its just slow..  ( couple of seconds per part ).. but hey it will be $300!!

WTF? My 15yr old P&P visioning is almost instantaneous on a Pentium MMX 233MHz. With the GPU goodies that a Pi has ... that is criminal!

The SoC on the Pi is optimised to be a setup box. So it is just good at decoding video. Not processing it. The CPU in the Pi isn't very beefy so I'm not surprised it is slow. The algorithm on your old P&P is probably better optimised as well.

It has a (I think) 700-odd MHZ ARM, so plenty of horsepower for simple image recognition tasks.
Video decoding is mostly dome by dedicated hardware

[mikeselectricstuff]

Just wondering: is there a universal programming language to control P&P machines, such as the G-codes for CNC machines, or is it every manufacturer has its own standard?

Not really because it is about far, far more than just moving stuff around - you need to integrate part library management, feeder allocation, vision parameters, component height fiducial data etc. in a way that makes sense for the way it needs to be used efficiently.

[Kjelt]

Thats a shame, it would make things much easier.

[sweesiong78]

   Its amazing that a raspberry PI can do the Vision.  It does it, its just slow..  ( couple of seconds per part ).. but hey it will be $300!!

WTF? My 15yr old P&P visioning is almost instantaneous on a Pentium MMX 233MHz. With the GPU goodies that a Pi has ... that is criminal!

The SoC on the Pi is optimised to be a setup box. So it is just good at decoding video. Not processing it. The CPU in the Pi isn't very beefy so I'm not surprised it is slow. The algorithm on your old P&P is probably better optimised as well.

I think the reason its slow is because none of the developers have experience in machine vision coding, from their own website:

"custom computer vision software to ensure that the parts are being placed correctly.  Computer vision is really hard.  Especially if you're not a math major."

there's alot of machine vision  and image segmentation algorithms published academic journals, but actually understanding the algorithm is hard, coding it efficiently is also hard. Im sure there are graduate students and professors out there working on this subject who can implement this in a trivial manner for them, but they are probably too busy with their research and probably also dont want to give away their IP for free.

[mikeselectricstuff]

   Its amazing that a raspberry PI can do the Vision.  It does it, its just slow..  ( couple of seconds per part ).. but hey it will be $300!!

WTF? My 15yr old P&P visioning is almost instantaneous on a Pentium MMX 233MHz. With the GPU goodies that a Pi has ... that is criminal!

The SoC on the Pi is optimised to be a setup box. So it is just good at decoding video. Not processing it. The CPU in the Pi isn't very beefy so I'm not surprised it is slow. The algorithm on your old P&P is probably better optimised as well.

I think the reason its slow is because none of the developers have experience in machine vision coding, from their own website:

"custom computer vision software to ensure that the parts are being placed correctly.  Computer vision is really hard.  Especially if you're not a math major."

Vision in a pick/place is far from hard as vision goes. Once you have the lighting set right, it's little more than thresholding and boundary detection. All you need to know is X/Y centroid position, size & rotation. 
I don't know much about OpenCV but I'd be very surprised if it couldn't do most if not all of the job.

[sweesiong78]

Vision in a pick/place is far from hard as vision goes. Once you have the lighting set right, it's little more than thresholding and boundary detection. All you need to know is X/Y centroid position, size & rotation. 
I don't know much about OpenCV but I'd be very surprised if it couldn't do most if not all of the job.

The 'thresholding' part is the tricky bit. I've quickly scanned FireSight which is the firepick implementation based on OpenCV and I see they have various feature detection routines like Hough transforms and a hole feature detector.
https://github.com/firepick1/FireSight/wiki

They could probably optimise their pipeline better for speed, and also they use a JSON pipeline management implementation, not sure if thats the standard or fastest way to do embedded hardware programming.

[nctnico]

   Its amazing that a raspberry PI can do the Vision.  It does it, its just slow..  ( couple of seconds per part ).. but hey it will be $300!!

WTF? My 15yr old P&P visioning is almost instantaneous on a Pentium MMX 233MHz. With the GPU goodies that a Pi has ... that is criminal!

The SoC on the Pi is optimised to be a setup box. So it is just good at decoding video. Not processing it. The CPU in the Pi isn't very beefy so I'm not surprised it is slow. The algorithm on your old P&P is probably better optimised as well.

I think the reason its slow is because none of the developers have experience in machine vision coding, from their own website:

"custom computer vision software to ensure that the parts are being placed correctly.  Computer vision is really hard.  Especially if you're not a math major."

Vision in a pick/place is far from hard as vision goes. Once you have the lighting set right, it's little more than thresholding and boundary detection. All you need to know is X/Y centroid position, size & rotation. 
I don't know much about OpenCV but I'd be very surprised if it couldn't do most if not all of the job.

I have done some projects with OpenCV. Thresholding etc is not easy to get right and almost never a solution in itself. In one of the projects I used a combination of methods to detect objects based on faint color boundaries.

[coppice]

I have done some projects with OpenCV. Thresholding etc is not easy to get right and almost never a solution in itself. In one of the projects I used a combination of methods to detect objects based on faint color boundaries.

Were you using OpenCV in something like a pick and place machine? In an arbitrary environment vision can be tough, because you have to live with arbitrary conditions. In a pick and place machine you control the lighting, you control the background, and the colours of the objects lie within a reasonably well defined set. That massively constrains the problem. Pick and place machines and chip bonding machines are very much designed to reduce the vision problem to some simple - in fact to something simple enough to run on a small box in the 1980s.

[a210210200]

Computer vision in Pick and Place is an easy computer vision task the environment is very controllable. The chips are high contrast objects, illumination is controlled, distances are very controlled, registration marks are a basically a given, pads are very visible and so on. Even the parts all should look very similar and you don't normally have squishy objects that try to run away or eat other objects in the system while your imaging.

An example of a hard computer vision problem is doing automatic classification in biological images which can look different every day. It is like asking a computer to circle a flat circle in an image and then asking the computer to then circle a cancer/malaria/... cell out of a billion plus cells based on morphology alone or count the number of birds in a image of millions of them in a natural landscape with the sun at a low angle and tons of other wild life/trees/...

A pick and place machine just does template matching and orientation detection on golden templates. Template based systems don't work very well outside a well controlled environment. Not only that the vision system has a ideal copy of the board as a map to guide it very few vision systems in real life uncontrolled situations have that luxury.

[sunnyhighway]

What I don't understand is why camera's are used in the first place. I mean, they are not easy to calibrate and the mount is very prone to vibrations and ambient light also poses a problem.

Why not simply use a simple flatbed scanner, get rid of the casing and glass plate, slap some wide soft brushes on the sides of the sensor to get rid of the influence of the ambient light and then mount it upside-down in the P&P machine.
Because the distance between the sensor and the pcbs is very small it is also very tolerant to vibrations.

Pro's:
 - Low price
 - Vibration tolerant
 - High resolution
 - Easy to calibrate

Con's:
 - ?

[mikeselectricstuff]

Why not simply use a simple flatbed scanner, get rid of the casing and glass plate, slap some wide soft brushes on the sides of the sensor to get rid of the influence of the ambient light and then mount it upside-down in the P&P machine.
Because the distance between the sensor and the pcbs is very small it is also very tolerant to vibrations.

For what? You are looking at parts, not the PCB.

[nctnico]

I have done some projects with OpenCV. Thresholding etc is not easy to get right and almost never a solution in itself. In one of the projects I used a combination of methods to detect objects based on faint color boundaries.

Were you using OpenCV in something like a pick and place machine? In an arbitrary environment vision can be tough, because you have to live with arbitrary conditions. In a pick and place machine you control the lighting, you control the background, and the colours of the objects lie within a reasonably well defined set. That massively constrains the problem.

Not so much. What looks like an equally lit surface to our eyes can easely have a 3 to 1 difference in luminance depending on where you are. Daylight and lamps with a weird spectrum can also be a nuisance.

[sunnyhighway]

Why not simply use a simple flatbed scanner, get rid of the casing and glass plate, slap some wide soft brushes on the sides of the sensor to get rid of the influence of the ambient light and then mount it upside-down in the P&P machine.
Because the distance between the sensor and the pcbs is very small it is also very tolerant to vibrations.

For what? You are looking at parts, not the PCB.

I was indeed talking about a cheaper and more precise way of determining the exact location of the pads on the pcb. Because it is the added position/orientation error of both the pcbs and the component that counts, the exact position/orientation of the component before it gets picked up can be less accurate in order to produce the same end result.

[mikeselectricstuff]

Why not simply use a simple flatbed scanner, get rid of the casing and glass plate, slap some wide soft brushes on the sides of the sensor to get rid of the influence of the ambient light and then mount it upside-down in the P&P machine.
Because the distance between the sensor and the pcbs is very small it is also very tolerant to vibrations.

For what? You are looking at parts, not the PCB.

I was indeed talking about a cheaper and more precise way of determining the exact location of the pads on the pcb. Because it is the added position/orientation error of both the pcbs and the component that counts, the exact position/orientation of the component before it gets picked up can be less accurate in order to produce the same end result.

Visioning the PCB is no use as it will have paste on it

[rx8pilot]

Fiducials. They don't get paste and that is what they are for.

[mikeselectricstuff]

Fiducials. They don't get paste and that is what they are for.

And why would you dick around with scanner parts when a camera can easily do the job perfectly well?
Spotting fids is even easier than measuring component offsets

[rx8pilot]

I will agree with that. The nice thing about a line scanner head is that it has little to no distortion compared to a camera single sensor/lens combo.

I am definitely in the camp that understands that this is not a real challenge, even for very modest hardware. In 1990, I was coding software to capture/analyze 3D space for motion analysis. Similar to what animators use today for character movement capture. Reversing geometric aberrations was not particularly difficult and we were scanning an entire 3D room with 286 computers at our disposal.

Having control of lighting and using a bandpass optical filter really limits what you are needing to process. Using a line scanner is cheaper and simpler than a camera and would not require any geometry correction. In the case of a P&P, the geometry errors are fixed and predictable, so that is not much of a reason to re-invent the wheel.

[mikeselectricstuff]

Having control of lighting and using a bandpass optical filter really limits what you are needing to process. Using a line scanner is cheaper and simpler than a camera

Nonsense.

[rx8pilot]

ok.

[a210210200]

I will agree with that. The nice thing about a line scanner head is that it has little to no distortion compared to a camera single sensor/lens combo.

I am definitely in the camp that understands that this is not a real challenge, even for very modest hardware. In 1990, I was coding software to capture/analyze 3D space for motion analysis. Similar to what animators use today for character movement capture. Reversing geometric aberrations was not particularly difficult and we were scanning an entire 3D room with 286 computers at our disposal.

Having control of lighting and using a bandpass optical filter really limits what you are needing to process. Using a line scanner is cheaper and simpler than a camera and would not require any geometry correction. In the case of a P&P, the geometry errors are fixed and predictable, so that is not much of a reason to re-invent the wheel.

Line scanners are complicated and to get a distortion free image you need a distortion free sweep across the plate and if you want to do it at both a high resolution and high speed on a rapidly moving machine your going to get mechanical distortion in the scanned images.

An optical camera can be made distortion free using very simple calibration grids and in some cases you can just spend the money and get basically no distortion optics.

A line scanner head still needs optics and try getting a cylindrical lens for your required working distance.

Another fatal problem with a line scanner is that if your processing a panel your going to need a massive line scan head to do it or cover it in multiple passes and have enough accuracy to stitch the image back together.

A line scanner is going to be more vulnerable to mechanical disturbances because the motion has to be controlled for the scan to work properly vs. a fixed camera which is not only going to be more compact is going to be mechanically one piece so it can shake all it wants but the pixels won't move relative to each other and with a fast enough exposure you can negate any vibration (Which both cameras will still have to deal with)

Also because your using a line scanner you need a similar uniform illumination system to follow it around, this is going to add a ton of bulk to the system vs a ring light around the lens of a regular camera.

Also telecentric lenses are super neat lenses.

[a210210200]

Fiducials. They don't get paste and that is what they are for.

And why would you dick around with scanner parts when a camera can easily do the job perfectly well?
Spotting fids is even easier than measuring component offsets

Using reference markers and scanning the entire board doesn't make any sense once you got the references in you have the entire board lined up and periodic checks for alignment are easier than scanning the whole board again.

Area scan cameras are easy to get and use and work fine for pick and place. Line scan cameras are meant for constant velocity belt inspection systems not XY pick and place.

[sweesiong78]

so with all that said, the firepick is slow at vision because of crappy programming with massive software overhead?

[rx8pilot]

That all makes sense. To clarify, my project from the 90's and all of my subsequent experience with single and stereoscopic cameras is with a traditional full field sensors and lenses. I have worked primarily in the entertainment industry, but got started in industrial imaging. My post was written in a way that may look like I was using line scanners which I have no experience with.

Measuring and correcting for geometric distortions is what we had to do on our 3D capture systems to capture people [athletes] moving. All of the corrections were post-processed but it was near real-time. I have also designed and built 360 camera systems for the military which of course required piles of corrections to make usable images. The good news is that we were able to get all that in real time via a single GPU - 8 streams of RGB 1920x1080. The final product had to be geometrically accurate because a separate system would composite 3D digital elements like people and vehicles onto our raw images. A training simulator system...


Anyway...
With today's tech, a global shutter monochrome camera is super fast and easy to interface. I have done a number of mirror and prism systems that could allow a single camera to do both top and bottom which would be nice. Maybe that is already how it's done with other machines.

[mrpackethead]

Firepick is $300 bits and some people hobby.   So go easy on it.

I'm looking very favourable at the MV10.