behavior around traffic circles

[sbma44]

Seeing some unexpectedly jagged traces when I simulate Dupont Circle in DC using the car profile.

The noise simulation seems like a reasonable approach to the problem, though not a particularly realistic one (looking at some actual telemetry I see a lot more parallelism than random jitter, perhaps due to either environmental RF echoes or the tendency of the location services black box trying to smooth its output by keeping heading relatively constant).

Assuming realism isn't a particular goal, I'll throw out two possible changes. Are there other ways of tackling this?

• having the location error as a cumulative vector that is adjusted at each step along a trip rather than a random error added at each step
• multiplying the error by a function of between-node distance to diminish its effect around circles

I'm also curious why the heading for the noise is normally distributed.

Happy to take a crack at a PR if any of this sounds like a good idea.

[morganherlocker]

I've been thinking about this issue and have a couple ideas. I believe that overall the GPS noise function is exaggerating the error distribution beyond realistic parameters per trace. In the real world, we do see a spread pretty close to what is simulated today, however, traces individually tend to have a higher relative accuracy. This is due to each device having higher relative accuracy than absolute accuracy. We can simulate this by giving each device a random drift bias in a randomly chosen direction, while simultaneously tightening the distribution of generic heading drift. This should give us less "squiggles" while retaining a nice spread across the road (we don't want data to be too "on the road").

multiplying the error by a function of between-node distance to diminish its effect around circles

Looking at your screenshots, I suspect there may be some additional interpolation issues when dealing with curves. The straightaways look far to clean by comparison, which is suspicious, since there are no intentional model decisions that would cause this. I'll do some more digging on this while I implement urban canyon noise.