[michaelt]

> keep in mind that car navigation will snap your position to roads.

Right. I chose driving as an example because someone navigating a city on foot could conceivably do it with low-precision signals like wifi. They can stop and wait for a better signal, or carefully study the map comparing it to street signs; a fix that puts them on the right block and that only updates once or twice per block is adequate, if not good.

Whereas turn-by-turn navigation is able to tell the driver "take the next right" 3 seconds before the turn - which shows the signal is being tracked precisely and regularly.

> For a better comparison, try e.g. Google Maps in pedestrian or bicycle navigation mode.

If you're into sports, take a look on Strava at people doing things like big city marathons: https://www.strava.com/activities/8078299417 https://www.strava.com/activities/51910242 where of course the positions aren't snapped to roads.

You'll see minor hiccups from time to time - but despite navigating between many tall buildings, you won't see errors of hundreds of miles.

[icegreentea2]

GPS units can also include INS or INS like logic systems to prevent those types of jumping around. It can be difficult to tell from an application output just how processed your "dot" is.

[michaelt]

It's actually fairly easy, for things like running GPSes.

You know the track isn't snapped to a map. And you know it's not built into a car, so no odometer or rate gyro.

Then just look at the trace as the sportsperson goes around a corner. Does the trace overshoot, then backtrack? That means the kalman filter is heavily smoothing with momentum. Does it produce a gently rounded corner that goes through a building? That means there's a moving average.

However if the traces display sharp corners when appropriate - then the signal is not being excessively smoothed.

(And I can assure you, even if a GPS watch has enough motion sensing to count steps, the swinging of arms and bouncing of feet mean wrist acceleration is noisy enough you've got no chance of detecting a runner turning a corner)

[icegreentea2]

Right, I was mostly thinking about robustness against transient jumps due to transient multipath reflection as you navigate around an urban environment. If you know that your strapped to a walker/runner, you know that you can't travel 100m in ~3-4 seconds.