0

What is the best strategy for determining a user's location from a series of GPS fixes, once they are considered not to be moving?

When tracking a user, if they should stop moving there will subsequently be a sequence of fixes roughly in the same location.

If possible I would like to not just use the last fix, but also take into consideration previous fixes so as to calculate a more accurate position for them.

Factors that I would have thought need to be considered:

  • The best way to determine a user is stationary (from experience speed from the GPS fixes is not sufficiently reliable)
  • Each fix has an accuracy, how can this be factored in?
  • Are there well established algorithms/libraries that could be used?

Any suggestions greatly appreciated

Has QUIT--Anony-Mousse
  • 76,138
  • 12
  • 138
  • 194
Mark Melling
  • 1,562
  • 14
  • 15

1 Answers1

0

What's wrong with taking the average?

If you want to take uncertainty into account, use a weighted average. Or a trimmed one, discarding those measurements that deviate most.

But it is a known fact that other factors such as reflections on buildings, can have a much larger effect on the accuracy.

Nevertheless, this is not so much a programming question, but one that needs GPS expertise. You are better off reading expert literature than asking random internet users for their opinion.

Has QUIT--Anony-Mousse
  • 76,138
  • 12
  • 138
  • 194
  • There may be nothing wrong, with taking an average or weighted average, although I would say you still need a way to determine which fixes to include (they were moving prior to stopping in one location), some clustering algorithm perhaps? I'm not convinced that GPS expertise really helps in this case, I would say that this 'expertise' is encapsulated in the accuracy figure provided in a fix. I wasn't sure if this was the right site to post this to, possibly Data Science might be more appropriate? – Mark Melling Nov 13 '18 at 11:28
  • No, migrating the question to GIS is more reasonable. Clustering is not reliable enough. K-means will split a very obvious uniform distribution right in the middle. Don't assume clustering can magically solve whatever you toss at it - it won't. Instead. It is fragile and needs to be carefully guided. So you may as well just do the median latitude and longitude, remove the farthest 10%, and do the resulting trimmed mean. It's just that all the GPS experts will tell you that your accuracy probably does *not* improve by using more fixes. Because that reduces only variance, not bias. – Has QUIT--Anony-Mousse Nov 13 '18 at 22:08
  • Thanks, I'll do some experimenting. – Mark Melling Nov 14 '18 at 16:16