Using vision to detect a small part in an area of interest

What is the smallest part the a depth camera could detect? I am thinking of using vision to detect small parts bouncing off a track into an area of interest.

It’s going to be hardware specific. Cameras for microelectronics assembly would be able to pickup components that are barely visible to the naked eye. Other constraints to consider end up being cost of camera, field of view, mounting location and position, etc.

I’ve seen this with anomaly detection where the “area of interest” is brightly backlit through white acrylic or something and then part detection is based on any dark spot detected against uniform background.

Hi Russ, thanks for the reply. I was planning to use the Depth Camera specified on the website, the Intel RealSense D415 and can see specs for depth and field of view but not size detected. However, I believe we have one of these in one of our offices so I intend to set up a trial to test it. we’ll take into account your comments on the background colour and backlighting.

@nweft01 For very small parts (on the millimeter scale) we recommend getting a specialized camera with zoom-in optics, rather than use the D415.
Since the D415 is meant for range (depth) measurements in the ~1m volume, it won’t be able to focus properly if you bring it very up close to the part, although I did have success with the camera at ~30cm (12") from the part and it focused just fine. At closer than 12" I’m certainly seeing degradation in terms of focus.

This is roughly what the D415 can do at 12"

This is D415 at ~6"

And at 3" it can’t focus any more:

1 Like

Hi Russ, I’m actually just trying to detect the presence of a part in the region of interest. For example a small disc or tablet moving into the region of interest circa 11mm in diameter or 22x9mm, these parts could have a depth of about 6mm. If the part moves into the area and passes through could it be detected or would it only be detected if it moved into the area and stayed there?
Best regards, Neil

I had to phone a friend to answer your specific question, so stay tuned for that.

In the meantime, for vision more generally you have your hardware specs but also the image processing specs that will determine the best approach to building your setup. Your first question about minimum detectable size is a matter of hardware, as is refresh rate (mostly). If the target object crossed the region of interest but entered and exited faster than the refresh rate for your image processing, you could completely miss the object even though it was large enough to detect.

As for whether the part has to stay in the region of interest, that’s a matter of configuration. The “detection” is basically an automatic analysis comparing images against each other. You can detect a part by having a static background and flagging any image that doesn’t look like the static background, which will generically detect any part and requires relatively little processing effort. You could also detect a part by searching images for something that looks like a target image. For the second case, think of a bin full of loose parts and a robot with vision camera that can identify the boundaries of a part.

Also, I’m not sure depth is a required dimension for what you’re describing. The stereoscopic image can help you get to dimensions, but flagging any anomaly in the region of interest seems like it would be sufficient even if no dimensions of the anomalous object are known.

Hi Russ, many thanks for the comprehensive reply. you’ve given me lots to consider. I think I’ll set up a small trial and see how it works. best regards, Neil

Shoot me a dm or email if you want to set up some time to discuss further