The project uses a couple of different approaches and algorithms.
There’s two general approaches to calculate positions from signal sources, without the cooperation of the signal sources or mobile networks.
Determine the location of signal sources from observations and then compare / trilaterate user locations.
Generate signal fingerprints for a fine-grained grid of the world. Find the best match for an observed fingerprint.
The second approach has much better accuracy, but relies on more available and constantly updated data. For most of the world this approach is not practical, so we currently focus on the first approach.
In theory one could use signal strength data to infer a distance measure: the further a device is away from the signal source, the weaker the signal should get.
Unfortunately the signal strength is more dependent on the device type, how a user holds a device, and changing environmental factors like trucks in the way. Even worse, modern networks adjust their signal strength to the number of devices inside their reception area. This makes this data highly unreliable while looking up a user’s position via a single reading.
In aggregate, over many data points this information can still be valuable in determining the actual position of the signal source. While observing multiple signals at once, their relative strengths can also be used, as this keeps some of the changing factors constant like the device type.
One other approach is using time of flight data as a distance metric. While there are some reflection and multipath problems, it’s a much more accurate distance predictor. Fine grained enough timing data is unfortunately almost never available to the application or operating system layer in client devices. Some LTE networks and really modern WiFi networks with support for 802.11v are the rare exception to this. These are so rare that we currently ignore timing data.