The camera doesn't know what the most important parts of the image are, and therefore what you would like to have in focus. Traditionally autofocus systems have simply selected the closest object near the centre of the frame and focussed on that. This can easily fail in the presence of irrelevant foreground details.

Face detection makes the assumption that if the image contains people, you would probably like to have their faces in focus, so by detecting the location of faces the autofocus system can be more accurately instructed.

In order to for such a scheme to work you must have some way of reading the image data continuously and the ability to focus on an arbitrary point in the image. This essentially means you need live view and contrast detect AF for face recognition. This is one of the reasons the feature was slow coming to DSLRs, contrast detect is still slower than phase detect and is thus not the primary way of focussing with DSLRs. Another factor is that DSLR users tend extert slightly more control over the camera's operation and prefer to select AF points manually.

The really interesting development is that the Canon 1DX has a metering sensor which is high enough resolution to detect faces, so it can use face detection in conjunction with phase detect AF.