Some things that I would adapt to your system
- Don't use keypoint matching for the meter part. I guess that will have a small accuracy over different meters.
- Like suggested by you, focus on distinct visual properties of the meter.
That being said try a colour base approach that does
- Translate image to HSV colour space (easier to seperate colors)
- Retrieve the red regions only.
- Find the largest blob after some erosion, dilation to clean up noise
- Find the bounding rectangle of that blob
- Now double the width from [0, width] to [-width, width], maintaining the height, because that is again a specific property of the meter that can be abused
That will get you pretty far!
