Hey Martin,
Happy to help 
I think maybe the recognition could be improved so your Pi can handle the program. I’ve never used a Movidius stick, what are the changed needed on the program to use it?
Have you tried something like haar-cascades for face recognition that should be much less demanding? If not, give this a try and if you think it will be suitable we can then adapt some code for the PWM signals. Face Detection in 2 Minutes using OpenCV & Python | by Adarsh Menon | Towards Data Science
With the SimpleFOC library you can disable the motors and this could simulate sleep mode you’re talking about. I’m not sure how important this is to you.
I disagree about the optical encoder being better for precise positioning. I’ve haven’t done a side to side comparison, but I would imagine they perform very similar (comparing a same resolution magnetic encoder to an optical one). Personally, I prefer magnetic encoders because they are cheap, compact, very easy to use and have been very precise for my applications. Furthermore, most magnetic encoders are absolute so as soon as you power the system you know where you’re at.