Case Studies
Machine Vision Measurement and Positioning
Our own vision software measures part dimensions and angle, then hands four-axis coordinates to the arm.
What we looked at first
The customer’s problem was not accuracy. It was speed. Operators measured each piece with calipers and then keyed the numbers into a system — two chances to introduce an error.
After looking at the parts we judged vision to be workable, and the reason was specific: contrast against the background was strong enough, and the range of placement angles was bounded. Without those two conditions, a vision approach turns into endless lighting and parameter work, and we would have advised against it.
Hardware
A Basler ace 2 series GigE industrial camera, monochrome sensor, with a C-mount fixed-focal lens, mounted on a perforated aluminium bracket beside an EPSON SCARA arm in a fixed overhead position.
Parts sit on a dark matte backing plate. That is not incidental. High contrast separates the part edges cleanly in the image, and it is the first thing that determines whether vision measurement holds up.
The software is ours
The measurement application in the footage is one we wrote. It runs on macOS and talks to the equipment over TCP.
Visible in the video:
- Area filtering Minimum and maximum area thresholds reject noise and anything that is not the target
- Dimensional measurement Detected parts are annotated in millimetres; the frame shows 4.3 mm, 2.7 mm and 2.3 mm
- Orientation The part’s rotation is computed; the frame reads 9.03 degrees
- Four-axis coordinates X, Y, Z and U. U is the SCARA’s rotational axis; without it, the arm has no way to know how far to turn to meet a part lying at an angle
- Parameter save and load, image rotation and mirroring, measurement logging
Why write it ourselves
Off-the-shelf vision packages are mostly closed. When they have to meet a customer’s existing line, what blocks you is rarely the algorithm — it is licensing and extensibility.
Writing our own means the measurement logic, the protocol and the data format all follow the site rather than the other way round, and machine learning can be added where it helps.
More detail
Measurement accuracy, cycle time and the range of parts we can handle are assessed against the actual part.