* Receivers in the metal-reinforced concrete canyons of a big city suffer from multi-path reflections of buildings, leading to bigger uncertainty. You can improve those signals with a (still pretty uncertain) rural reference. This reference typically also sees low angle satellites, and it receives all signals without reflections.
* You can improve the uncertainty of some of your references by running analysis on it's position history and compare it with (highly certain) map data. One of your rural signals is moving at 78 mph? High likelihood to be in the left lane on that highway. Moved with 30mph, now suddenly stopped? Waiting at that specific traffic light, probably first car.
With crosschecks between enough references, some averaging and smoothing, you can produce a usable reference. Add a couple of absolute references (no reason for the company to not have a fixed antenna in the 50 largest cities each), and you're competing with current commercial reference services.
It's also worth noting, that for many of the uncertainties of the GPS signal (satellite clock jitter, atmospheric conditions), you don't need to update your position delta every second. It's already an improvement if you get a good reference update every minute or so.
The whole reason technologies like DGPS work is because receivers in roughly the same location will have roughly the same offset. If you average out the stationary receivers, you will be left with a shared common drift - which can be corrected for. Slow-moving receivers probably don't affect the average too much, and you'd be able to filter them out over time as they'd have movement beyond the shared common drift.
Devices will enter the stationary set as they stop moving, and move out of it as they start moving again. Even someone stopped at a traffic light for a few dozen seconds might be able to contribute to it.