Hey deohboeh,
Welcome to tech-artists.org and the wonderful world of tech-art. 
Facial animation can be done in a manner of ways, but one of the most common setups these days is to have the face driven by joints for most movement, with blendshapes for more extreme or sculpted emotions or a combination of blendshapes to hold volume in places like the cheeks, etc… If you’re looking at that, would be good to watch Judd Simantov’s TLOU rigging video to see a visual demonstration of this system. That can be viewed here:
Blendshapes can be quiet intensive on an engine, although those barriers are getting broken down more and more for next-gen (current-gen?). So the most efficient way is to have a solid base of joints which move the face via skinning, and blendshapes only added on when and as required. However, all these variables are down to the level of detail you need, the limitations/constraints of the game you’re working on and engine limitations. I’m not sure how UDK handles blendshapes, so that’s something to look into too.
Your best bet is to sculpt a neutral face shape, eyelids relaxed, mouth partially open, etc… This is just because making changes from neutral shapes is generally easier, especially hitting extremes on both sides. However, this isn’t always the case as you may need to make special cases or modify stuff for particular reasons. However, it’s a great place to start. You’ll also want to really nail the topology. There is a lot on polycount’s wiki on facial topology, which is a great place to start. Good facial topology can make or break a face, so definitely make sure you put the effort in here.
http://wiki.polycount.com/FaceTopology
15k-21k would be tris. This is because real-time models are almost always triangulated in-engine. What you do, making wrinkle maps and such, is really up to you. If you think you’re character or game sequence can support that or you think you need that to get the desired shapes, then feel free to do it. There will be a lot of information online regarding this, just search for “blending facial maps” or something similar. You can definitely have that in your range of emotion, mocap is just a way of capturing human movement and is usually transferred on to rig anyway as only facial scans or systems like LA Noire’s capturing technology really get that level of detail.
Hope this helps to answer your initial questions. Feel free to ask more or for clarification if you need.