yes, I agree with your analysis. One aspect I'd like to emphasize, though, is that with 'intelligence' we need distribution, i.e., the sensors need to talk to each other.
Now for the discussion on the use case: You're perfectly right, if we predict we are operating under probabilities. First assumption is that the person continues moving and does not simply stop.
Sometimes these probabilities are very high, for example, when a person walks down a hallway or opens a door. Sometimes these probabilities are lower, if a person can walk right or left for instance.
Our use case is that we want (with some threshold) support continuous lightening on some price of energy consumption. Typically light sources (incl. the sensors) are not densely covering the area. In our university for example, I have to deeply enter a lecture hall until a sensor detects me and switches on the light. This is really unpleasant at complete darkness. It is also inconvenient, if people walk down a path and lights turn on too late. Then a steady change in brightness is affecting the vision.
However, there is some tunable parameter (energy versus comfort) and there is learning potential (i.e., how often did a prediction succeed?). I guess this gives a rather rich field of building a solution that covers quite a number of aspects and focuses on the really exciting topic of a distributed intelligent system.
Does this help?