Significant drift when recording using Run+

[danielittlewood0]

Hi, I am having issues using the Run+ app. I can see lots of discussion and several apps relating to getting a GPS fix. I don't think this is my problem - when I head outside (or put my arm out the window) I get a green GPS label after just a few seconds.

However when I go out to record, I am seeing a big difference between what I record on my phone and on my watch. In particular, when standing completely still the distance will increment, sometimes by as much as 2m per second. I have tried the Recorder app set to record every second, and every 10 seconds (the default), and I see drift in both cases. I am not sure whether it is better or worse in either case.

I tried using the GPS Info app, to see what that is reporting. I watched it for about 5 minutes while standing completely still, and saw the following decimals in my lat/lon (just the decimal parts so I don't dox myself).

max lat: .354179
min lat: .354149
max lon: .230940
min lon: .230905

in the extreme case this seems to mean the GPS uncertainty can be as high as 4m. This was with a fix of 9 satellites. My altitude also seemed to bounce between 9m and 15m.

I found this old thread, some people mention that the drift should be worse if you record more often (I guess you want the uncertainty to be a smaller proportion of the true measurement). I'm not sure if anything has changed since then. https://github.com/orgs/espruino/discussions/7298

[danielittlewood0]

On reflection, maybe this is all kind of obvious. Wikipedia suggests typical GPS uncertainty is around a meter. So even the chip in my phone which I trust should expect to drift as much as a kilometer during a half-hour run. In fact I just tried recording a non-activity with Strava on my phone, and while if I leave it alone it drifts extremely slowly, if I shake the device then it drifts very fast.

It seems like some heuristics based on the accelerometer could make distance tracking much more accurate. I tried reading the run+ code but can't immediately see where it calculates distances. Is that in Recorder perhaps?

[thyttan]

Yes much of that will be in the recorder app iIrc. Whether potential improvements to the heuristic should go there or somewhere else I'm not sure.

[danielittlewood0]

I found where the calculations happen, it's the exstats module: https://github.com/espruino/BangleApps/blob/5e21e41bb61f1058ef6a81fb8f32579e9e35f1f0/modules/exstats.js#L9 (see calcDistance and related functions).

I've loaded some recorder data to my laptop, I will do some experiments in python later and see if I can roll a more accurate distance function. There are a lot of potential options to try.

[gfwilliams]

Hi! Yes, this is something you unfortunately expect a little. Especially when you start as the GPS tries to give you its best guess initially, and then slowly hones in on the best value - your phone will have a big advantage as it can download the absolute latest GPS correction data and may also have had the GPS powered on for something else recently too.

As you found the exstats module is the one that's responsible for the run stats. recorder will just record the raw GPS data and then it'll be up to whatever reads that to use its own heuristics.

However you can't just 'smooth' the GPS data as it'll cause under-reading when you're going around corners (imagine what happens if you're running up a track, turning around, and going back)

There have been some discussions about adding a 'Kalman filter' to the data that comes from GPS, which I think would be a great idea. There was one in a previous running app years ago but it never seemed to work properly, but if you could get something working I think it'd help no end (and you could potentially use the step counter and stride length as an extra input for velocity).

Taking recorded info (best to use 1 sec per sample data as that's what exstats will work with) and doing some experiments as you're suggesting sounds like a great way forward. With this kind of thing it's often really hard to not make things worse if you're just experimenting - it's much better to have real data that you can try different approaches on.

The GPS also provides a 'confidence' value called HDOP which could be fed into a Kalman filter, but it's not generally very well-defined exactly what it is - so actually having in a bunch of data you can experiment with to get the right scaling value for HDOP would be perfect.