Notes on GPS accuracy problems

Owned by Diarmid Mackenzie
May 14, 2020
1 min read

Some notes from some folks in Intel research on GPS acuracy:

  • GPS has to train for 15-20 seconds before it gets decent accuracy (matches what we have seen)

  • Training can take much longer in bad weather conditions

  • GPS cold start uses a lot more battery than warm start - it’s much more efficient to keep GPS running than to stop & start it. (we think we know that keeping GPS runnning would use too much battery, but we’d been hoping we could use a lot less battery by using it for only 15 secs every 5 mins - we may end up disappointed on that).

  • GPS performs especially poorly in urban areas, especially “urban canyons”. Google Fused helps a lot here.

  • Self-reported accuracy is generally pretty reliable, but can be inaccurate in urban environments with GPS signals bouncing off glass.

  • Most of the Intel location research ended up being built on top of Google Fused, so that’s what they know best (and it performs pretty well, with lower battery usage).

We also discussed some ideas for fixing up GPS data on users' phones

  • Algorithms to spot clusters of points, and classify them as “visits” - might help us to eliminate noisy data points.

  • We can do this already during redaction for infected patients data - but this might be valuable for non-infected users.

  • This could involve user interaction: “it looks like you just visited X - can you confirm?”