Project ZeroAR

Well first off, I think funny hats are in order.

If you require funny hats, then tracking someones head becomes trivial.

A simple down looking camera could track the top of the persons head from below the projector.

I'm thinking a couple super simple QR codes could give roll/pitch/yaw.

Add 6 QR codes and you could track arms and wrists.

Strobe some colored LED's on the fingers of some gloves, and use two cameras.  Now you have Minority Report.

I think it might be simpler to adopt some BSP wielding game engine.  Last one I used was the unreal engine.  No idea of the state of develpment on any of these now.

As a game client, stream into the game server our real world inputs.  

The game map would need to be large enough at the start to represent whatever size physical battle field is being used.  

Simple height and shape mapping for buildings should probably be present to prevent AR bullets from firing through real-world obstacles (unless you have a rail gun).

I think the ideal playing field would be a park in need of some serious grooming.  :)  Basically flat or lightly wooded.

It would be the responsibility of the individual node to output as accurate of positional information it could.  This would stand on GPS and be augmented by magnetic, inertial, and gyroscopic sensor fusion.  They would serve as inputs to the game engine the user can not control.  The real-world angle of the machine is used when manually targeting an objective for example.

With accurate positional and vector information, generating a suitable overlay for the user in the game engine should be possible.  Direct projection of the exact shape/size of the AR represented opponent should probably be avoided to prevent artifacts at close range, but AR targeting of opponents you can't actually see yet would be cool.  See through walls upgrade acquired.

Localized effects such as a shield centered on the opponents position could be used though and the relative positional information of the opponents can be used for hit detection using the game engine.

Damage could be simulated on the possibly briefly illuminated AR limbs of an opponent and that damage could then be translated to partial loss of motion for the associated hydraulic control.