The Gottshalk reference above is probably the best one on OBB collision detection. That's where I learned about it from, anyway. I'll do my best to describe it here.
Let's start in 2D. The first thing you have to understand is how to project objects onto a line (an axis) in space. You can imagine this, visually, by placing a spotlight directly over the objects and watching where their "shadows" would fall on the line. (The virtual spotlight should be oriented so that it shines directly onto the line; its light rays should be perpendicular to the line) The shadows are the projection of these objects on the line.
If you don't know how to mathematically find the projection of one vector onto another, then you should take a course or read a book on vector math, dot and cross products, etc. I'll assume you know how to do this.
Anyway, imagine two boxes in quadrant I of a 2D space. Let's use the x-axis as our "line" for simplicity's sake. Now imagine the projection of these boxes onto the line. If the projections don't overlap (you have two distinct projections with some separating region between them), then the boxes cannot be colliding and you have found the "separating axis". If they were colliding in any way, then the shadows/projections would overlap.
Keep in mind that if the shadows/projections DO overlap, that doesn't mean that the boxes are colliding. It just means you can't tell yet. You have to try another axis. Try the y-axis. Still overlapping? How about the line formed by one of the box-edges? The Gottschalk paper describes the minimum number of such tests required to find a separating axis. If you exaust all of those tests and still can't find a separating axis, then the boxes must be colliding.
In 3D, the same analogy applies. Imagine the shadows of the boxes on some arbitrary plane (since we've added a dimension, we're looking for a separating plane now). If they don't overlap, then you know the boxes can't be touching each other. If they do, then you have to try another plane.
Ultimately, it turns out you have to try, at most, 15 such tests before you can be sure that the boxes are colliding. The paper describes a lot of simplification in the math that makes these tests a lot faster computationally. I also think it does a great job explaining it. You just need to slow down and take it one step at a time.