Just to be clear. I do think the accelerometer/hardware part would be very difficult. I also think the software side of that would be super hard which is why I'm guessing it would take over a year once you have the hardware. In fact because it is so hard I think the first way I would approach a solution would be with an ensemble machine learning solution leaning on random forests because they are quite fast and Random Forests is the ML algo I have had the most success with in the past, and then try and post process that. If that worked I would probably have users individually generate their forests during a calibration when the game starts.

I have 2 more ideas if that doesn't work well. Though I would probably end up combining all the ideas in another weighted ensemble on top.

The game part of this however is not really that complex and could be done on a regular schedule.