Both really. It depends what you're targeting. If it's "dog walking" then you check density for "dog walking" as opposed to "dog" and "walking". If it was "dog walking swindon" then you'd have a density increase for both "dog walking" and "dog walking swindon".
Likewise you'd naturally have an increased density on the words "swindon" and "dog" which on their own aren't hugely relevant. Nowadays they say keeping an eye on LSI (latent semantic indexing) is actually more important as it shows relevance between words and gives the more accurate impression of natural speech. (Just to confuse matters!)

EDIT: Just to clarify, that's the way I always calculated it, not to say it's how your client will!
