My assumption was, that reading a simple value of a gauge should be feasible with a Raspberry Pi or cheaper hardware, some software and little configuration. But it seems we are deep in AI here.
Training various needle positions is not possible since I cannot unplug the gauge an make image calibrations with specified analog values.
I'm still interested in a noninvasive solution for reading analog Gauges without installing sensors / additional hardware. Of course postprocessing is done with NodeRED! Ideas are welcome.
What a coincidence. That gauge looks pretty much like mine at home
I find your question fascinating, because like you say it is a "noninvasive" approach. So no changes required to the plumbing, no corrosion of the sensor, ...
For example this Github project might be interesting for you: actively maintained, lots of contributors, lots of stars, ...
I have not studied it into detail. E.g. I don't know whether you can train it for another gauge...
BTW it runs on a cheap esp32, and it can publish the numerical output to an MQTT server (see wiki). So you can easily integrate this into e.g. Node-RED.
It seems to me (see here) that you can do it all via a nice web interface (hosted on the esp32): capture a snapshot reference image from the cam, and draw your regions of interest...