Hi universalmaster,
I had very similar issues than you:
1.) The HFR values are "noise", so to get a reliable v-curve, I either had to use long focus-exopsures or averaging over many frames. This made autofocus very slow and not practical.
2.) When focusing with the full field, the different detection algorithms either detected to few stars, or to many "false" stars (noise, hot pixel).
3.) The the different focus algorithms move all very far out of the focus, so that eventually they don't recognize stars properly anymore. That gives completely wrong measurments for the HFR values at the edges of the v-curve.
All these issues prevented me from using autofocus automatically and without supervision.

I found a solution I am very happy with:
- Don't focus on your actual target, but focus on a bright star close to your target (really bright). (create a "fake" job in your scheduler, to slew to a bright star and focus, before slewing to your actual target).
- With a short focus-exposure time of 0.5 seconds, star detection "gradient", "Auto-select star" and "subframe", the algorithm reliably picks up the very bright star as the target to focus on. The short exposure time and the subframing means, that you can rapidly take many frames to get a more averaged (and less noise hfr measurments). (I average over 8 frames)
- The linear algorithm works best. Even if it moves far away from the minimum of the v-curve, the star is so bright that the algorithm still detects it reliably.
Because the v-curve is very clear and not noise at all, I can choose a tolerance of 0.5%.

With these settings, the automatic focuser works without intervention flawlessly and very quickly.

The only downside is the requirment to work with "fake" jobs in the scheduler, to slew to the bright focus-star of your choice.

Read More...