Edge 305: inaccurate record of elevation using 1 sec record

Post a question about the software or find answers here.

Edge 305: inaccurate record of elevation using 1 sec record

Postby CHnuschti » Sat Jul 25, 2009 6:20 am

Very inaccurate record of elevation when using 1-second record of the Garmin Edge 305. (Original thread)

Did analyse the behaviour of the "1-second-records". As known, in the Garmin Edge 305 you can either choose "smart record" or 1 second record.

The general conclusion of the analyse is, while HR and cadence are recorded very reliable, the elevation record is a desaster, with 1 second record you will get an oversized ascend/descend overall that usually is approx. 2 times as big as what you get when you compare with real elevation data from digital maps etc., it gets worser the more flat your ride is. While as far as I can tell, the elevation record (overall climb) with smart record matches very precisely with real data from digital maps (max. error 10%, usually 5%).

IMO there is also a simple explanation. If you look at the ORIGINAL records, you will notice that ALL the elevation gain/losses between the record points are always multiples of 0.48 meters (approx. 1.5 feet). So 0.48 meter obviously is the smallest difference of elevation that can be sensed by the barometric measurement. Considering you usually ride with 20-30km/h (approx. 13-20 mph), your distance per second (of recording) is 5 to 8 meters, so you would have to ride a grade of 6% or more (at this speed!) to reach this elevation step of 0.48m.

In your record, you will have tons of following entries with elevation differences + and then -0.48m or viceversa (or multiples of it). As it seems, the sensor is to unsteady in its sensing, fluctuates and eventually causes the inflated overall ascent/descent. Even when you are indeed climbing grades of 5%, you still will have negative gains.

The remarkable point is, how much the 1-second record worsens the overall climb/descent, compared with the almost "precise" smart record (which usually has an average of approx. 1 entry every 5 seconds).

Don't want to know what climb values the Forerunner 305 (which has not a barometric sensor) produces with the 1-second record :D.

regards

EDIT: BUT, there is to say, the observation applies only if you analyse directly the record (or export it into another application such as SportTracks). In the Training Center of garmin, you will have an ascent/descent that approximately matches the "reality". So obviously TC applies some smoothing and/or correction algorithms.

Additional note: These observations very likely are still valid also for new devices such as the Edge 705 etc., since its barometric sensor seems to be still the same as the one of the Edge 305.
Last edited by CHnuschti on Tue Jul 28, 2009 9:09 am, edited 1 time in total.
CHnuschti
 
Posts: 267
Joined: Tue Feb 13, 2007 3:28 am
Location: Switzerland

Postby pkan » Tue Jul 28, 2009 4:12 am

I guess the conclusion is that if you have elevation data quantized in steps as large as 0.48 meters, with some measurement noise added before quantization, then you need a bit of preprocessing before attempting to compute ascent/descent/grade, at least if 1 second recording is used. ST has smoothing settings that might do some good, but a better approach would probably be to come up with an algorithm that is adapted to the behavior of the Edge elevation data. Such preprocessing could be implemented in a plugin.

This is somewhat related to the problems of the distance calculation, where lateral noise (i.e., perpendicular to the direction of motion) in the GPS position data leads to the total distance always being overestimated by ST with 1 second recording, see http://www.zonefivesoftware.com/SportTracks/Forums/viewtopic.php?p=10370&highlight=#10370. Unlike elevation data, total distance is not affected by ST smoothing, since smoothing is applied to the computed distance track rather than to the raw GPS points.
pkan
Donated!
Donated!
 
Posts: 145
Joined: Fri May 12, 2006 2:26 pm
Location: Sweden

Postby CHnuschti » Tue Jul 28, 2009 9:15 am

Some remarks on this:

- These measuring steps of 0.48 meters are not a characteristic of the Garmin devices with barometric sensor alone. I also thought this first, but on a closer look on the records of an Edge 705, Forerunner (FR) 305/405 (FR are without barometric sensor) I determined they all measure elevation in these steps of 0.48m. I wouldn't be surprised if this applies for ALL Garmin devices able to record the elevation, to me it seems to be a general design.
So this "fluctuation" of elevation between altitude steps differing 0.48m in principle is present in all these devices. In devices with detection of elevation alone with GPS (Forerunner etc.), I guess this has little effect, since the fluctuaction of the GPS measurement itself IMO if far greater.

- Edge 305 and 705 devices have an own calculation of total ascent/descent WITHIN the device. Based on what I examined, IMO Garmin Training Center (GTC) just takes over these values from the units, it does not calculate them. In few cases I could notice a very small difference (promille) between tot. climb of GTC and within the unit, I suppose this is just a rounding issue or similar at takeover of the data from the device.
Opposite to this, FR have NOT the tot. ascent/descent display ("field") in the devices. Now it's unknown if this values is calculated within the device, IMO it is unlikely, so GTC in this case will calculate it.

- Be that as it may, it is remarkable that tot. climb given by devices and/or GTC does not differ very much, if recorded "smart" or at "1second". In opposite to ST which gives huge differences for these values depending of the 2 recordings types. So there is to admit the Garmin algorithm works remarkably well.

- As far as I remember, already a minimal smoothing setting in ST of "1" has a high impact on the tot. climb of 1-sec-records, it already wipes out a major part of the "error".

regards

Edit: have 2 Edge 305 here available ATM, will record simultaneously "smart" and "1sec" and do a comparison for the sake to the fun, ahem, science. ;)
CHnuschti
 
Posts: 267
Joined: Tue Feb 13, 2007 3:28 am
Location: Switzerland

Postby CHnuschti » Tue Jul 28, 2009 6:24 pm

Well, did a first test today. Edges were used simultaneously.
red = smart record
blue = 1-sec record
If you look at the overall picture, you would say it is an excellent correspondence for the elevation. In fact it is, what concerns the absolute elevation. But if you zoom in a lot you can see clearly the said toggling +/- 0.48 meters in the 1-sec record, which falsifies the overall ascent/descent so much.
Image Image

As for the "numbers", it does not give the thoroughly clear "picture" I had in mind from the past:
- Tot. ascent/descent is definitely different within the device and in Garmin Training Center (GTC), my statement in the previous post is definitely wrong.
- For averages it can be said that smart record is "enough", you don't get any relevant improvement of the "precision" with 1-sec record.
- For distance, also here differences are irrelevant, differences are <1%
- Tot. ascent/descent gives not a clear picture. Sure is, 1-sec record imported into ST gives a very oversized climb (unsmoothed). GTC handles this definitely better. On the other hand, setting elevation smoothing to 1 (its minimum) wipes out almost the entire "error" of the climb in the 1-sec record, while this setting does only affect the smart record moderately.
- As for the tot. ascent/descent from Swiss Map, I would give more credit to the values of the smart record. Average distance between record points is 40m here, while it is 7m in the 1-sec record. Elevation contours are 10 meters (5m when "flat"), in between there is an interpolation. I think this leads to additional, summing up errors in the 1-sec record.


The numbers (will add other rides in the next days):
Image


Testequipment:
Image


regards
Last edited by CHnuschti on Thu Jul 30, 2009 7:24 am, edited 2 times in total.
CHnuschti
 
Posts: 267
Joined: Tue Feb 13, 2007 3:28 am
Location: Switzerland

Postby texmurphy » Tue Jul 28, 2009 6:57 pm

Looks like elevation start has a greater variance then elevation end on the Overlay view (assume round trip to start).

Idea for subsequent tests would be to negate handlebar vs stem mount vibration / aerodynamic pressure effects by swapping Edge 1sec with Edge Smart positions.
texmurphy
Donated!
Donated!
 
Posts: 2128
Joined: Wed Jul 05, 2006 7:38 pm
Location: Maryland, USA

Postby admin » Tue Jul 28, 2009 7:01 pm

For what its worth the elevation is communicated by the device as a float (32 bits with 23 bits mantissa) which gives a precision of 6-7 digits.

I think an optimal scenario would be to record in the maximum resolution (1 second), then clean and compress the data after the fact during import. This process would have to be tailored to each GPS manufacturer, and possibly each device. It seems a lot easier to trust the device's "smart recording".

It seems for some tracks such as elevation there is really no sense in storing the data at 1 second intervals, however for others it might be nice. Do you know when you are smart recording if the tracks are storing data points at different intervals (such as heart rate by the second)?

A feature I found interesting on the Globalsat watch is the ability to set the recording interval - 1 second, 2 second, 3 second, etc.
admin
Site Admin
Site Admin
 
Posts: 3013
Joined: Tue Apr 05, 2005 8:52 pm
Location: USA

Postby CHnuschti » Thu Jul 30, 2009 8:00 am

admin wrote:Do you know when you are smart recording if the tracks are storing data points at different intervals (such as heart rate by the second)?

No. Exporting a *.tcx of the activity from GTC, you get repeating entries of this kind:
Code: Select all
          <Trackpoint>
            <Time>2009-07-29T17:01:12Z</Time>
            <Position>
              <LatitudeDegrees>47.4410943</LatitudeDegrees>
              <LongitudeDegrees>8.4326977</LongitudeDegrees>
            </Position>
            <AltitudeMeters>434.1726074</AltitudeMeters>
            <DistanceMeters>10520.1054688</DistanceMeters>
            <HeartRateBpm xsi:type="HeartRateInBeatsPerMinute_t">
              <Value>135</Value>
            </HeartRateBpm>
            <Cadence>102</Cadence>
            <SensorState>Absent</SensorState>
          </Trackpoint>

So smart record seems to be triggered to collect ALL available data at a particular record timepoint, probably there is an algorithm continously screening all values and then trigger the record entry of "everything" as soon one of them changes relevantly.


IMO 1-sec record does not make sense in general. All you get is maybe some more precise maximum values, but in general these are quite useless. What counts is the average IMO.


Did another ride. Units were swapped on their mounts. This ride was very flat, same scale as first ride is used in the pic.
Elevation of 1-sec record (red) was 10m off at start and end.

Previous statements remain valid here.
Note the ascent/descent in the devices. 1-sec record gives a notably lower value than the smart record. ???. Quite incoherent, one wonders what algorithm is used here by Garmin.
ST smoothing="1" affects the 1sec-record very much, it's maybe "over"-adjusting it.

regards

Image
Image

Image
CHnuschti
 
Posts: 267
Joined: Tue Feb 13, 2007 3:28 am
Location: Switzerland

Postby markw65 » Thu Jul 30, 2009 11:58 am

CHnuschti wrote:Note the ascent/descent in the devices. 1-sec record gives a notably lower value than the smart record. ???. Quite incoherent, one wonders what algorithm is used here by Garmin.


Note that the difference is only 6 or 7 meters.

From this article the arrival time of the gps signal has a resolution of about 10ns (in "modern" gps units). Which is about 3m.

But for elevation, the signal is (typically) arriving at a very low angle. So 3m of error in the distance from the satellite could translate into a much bigger error vertically.

My guess is that the two units internal clocks sync'd up a few ns apart, and as a result recorded (consistently) different altitudes, and that it had nothing to do with smart vs 1s recording.

An even simpler explanation (since I just remembered that this is a barometric unit!) - the gps was presumably only used to set the initial altitude (and was subject to the usual gps altitude issues), and then the barometric altimiter took over. So not surprising that whatever difference appeared at the start was maintained throughout...

Mark
markw65
Donated!
Donated!
 
Posts: 435
Joined: Fri May 09, 2008 4:35 pm
Location: San Jose, CA

Postby gerhard » Thu Jul 30, 2009 4:38 pm

admin wrote:I think an optimal scenario would be to record in the maximum resolution (1 second), then clean and compress the data after the fact during import. This process would have to be tailored to each GPS manufacturer, and possibly each device. It seems a lot easier to trust the device's "smart recording".


As the device has access to information like number of satellites and accuracy, the device can make a better job of filtering than a post processor (even if the device uses black magic).
Of course, an even better solution is if a post processor also had that information.

For non GPS tracks, the algorithm does not seem to be so smart, but I do not care myself about that.
gerhard
Contributor!
Contributor!
 
Posts: 1798
Joined: Mon Nov 27, 2006 7:07 pm
Location: Sweden

Postby CHnuschti » Fri Jul 31, 2009 2:21 pm

Well, did other two rides and will conclude the comparison with it.
Last one 20090731 was almost entirely in the woods (of decent densitiy and height 15-20m) with the mountain bike. I supposed the GPS signal would be poorer there and therefore maybe have an influence on the "numbers". This is not the case at all.

Beside that, no new findings. So conclusion is, distance is almost not affected by 1-sec record, nor the averages are. It's only the elevation sums which are considerably off. Maximum values also differ, but I wouldn't give too much credit on this.

Noticeable is, values in GTC and within the units differ relevantly. I could swear, in the past (older version of GTC and/or Edge firmware), these values were always +/-identical.

regards

Image

ImageImage

Image

Image
CHnuschti
 
Posts: 267
Joined: Tue Feb 13, 2007 3:28 am
Location: Switzerland

Postby CHnuschti » Fri Jul 31, 2009 2:24 pm

markw65 wrote:Note that the difference is only 6 or 7 meters.

Columns to be compared are E and F, I think you did read the wrong ones.

And, with respect, I think you're way off with your analysis. I don't think it's a question of unsynched units etc..
It is just the algorithm of Garmin which fails here in the very flat 20070729 to filter out the +/- 0.48m toggling. Sure Garmin is aware of this, but also has not the perfect algorithm for this elevation errors in the 1-sec record. Also ST smoothing=1 "over" corrects this here (tot. asc./desc. is lower than in the smart record), so it seems both used smoothing algorithms have something in common.

The "problem" accentuates the more flat a track is. I'm pretty sure, if you would make this comparison on a perfectly even track, you would get differences at their "best".

I guess it's difficult to find a suitable algorithm correcting the elevation working well in all cases, since the toggling +/- 0.48m occours somehow randomly. At least there is to remark the ST-algorithm works remarkably well in this case, as said a setting=1 wipes out the most of the error.

regards
CHnuschti
 
Posts: 267
Joined: Tue Feb 13, 2007 3:28 am
Location: Switzerland

Postby pkan » Fri Jul 31, 2009 3:55 pm

CHnuschti wrote:So conclusion is, distance is almost not affected by 1-sec record, nor the averages are.

Note, however, that this refers to biking. To first order, the overestimation of distance in ST due to lateral measurement noise (i.e., false zig-zagging) is ~1/(speed*(time between measurements))^2. At running speeds this error becomes significant with 1 second recording, in my experience typically 2-3%. This would correspond to <0.5% at biking speeds, which is consistent with your observations. For walking I'd expect ST to overestimate distance by >10% with 1 second recording.

With smart recording the distance error in ST from lateral measurement noise is reduced by a factor of at least 20, typically making it insignificant. But of course there may be other sources of error in the GPS data, which will also affect the calculated distance.

As for the distance track produced by the unit, I haven't noticed any difference between 1 second and smart recording, which seems reasonable since I'd expect them to be the same thing, only with the output from the internal filtering algorithms sampled at different rates. But as far as I know no details have been revealed by Garmin.
pkan
Donated!
Donated!
 
Posts: 145
Joined: Fri May 12, 2006 2:26 pm
Location: Sweden

Postby markw65 » Fri Jul 31, 2009 4:11 pm

CHnuschti wrote:
markw65 wrote:Note that the difference is only 6 or 7 meters.

Columns to be compared are E and F, I think you did read the wrong ones.


Sorry, I think I misread your comment "1-sec record gives a notably lower value than the smart record. ???. Quite incoherent"

I thought you were pointing out how odd it was that the /absolute/ altitudes were different (ie, the fact that the unit in smart record mode read a consistent 5/6 meters lower than the other) and wondering what kind of algorithm would do that.

And I stand by my explanation of that :-)

And, with respect, I think you're way off with your analysis. I don't think it's a question of unsynched units etc..
It is just the algorithm of Garmin which fails here in the very flat 20070729 to filter out the +/- 0.48m toggling. Sure Garmin is aware of this, but also has not the perfect algorithm for this elevation errors in the 1-sec record. Also ST smoothing=1 "over" corrects this here (tot. asc./desc. is lower than in the smart record), so it seems both used smoothing algorithms have something in common.

The "problem" accentuates the more flat a track is. I'm pretty sure, if you would make this comparison on a perfectly even track, you would get differences at their "best".


Right - I fully agree with all of this, and wasnt addressing it in any way in my previous post.

Mark
markw65
Donated!
Donated!
 
Posts: 435
Joined: Fri May 09, 2008 4:35 pm
Location: San Jose, CA

Postby CHnuschti » Sun Aug 02, 2009 4:00 pm

pkan wrote:For walking I'd expect ST to overestimate distance by >10% with 1 second recording.

Well, took note about your statements, but had my doubts, so I made a test walking (not jogging, just walking).
First part was completely in the open field, second part was exclusively in dense woods on a small footpath, where the sky was covered throughout.
Differences as you can see are equal in open field and <3% in dense woods where I suppose GPS was quite poor. So at least this test does not confirm any major influence of 1-sec-record on the distances.

Image

Image

Maybe for Forerunners, a test would give different results. IMO a relevant point is the position of the device. While on the bike (on bar/stem) IMO position is optimal, IMO for Forerunner this is not the case. I guess a compromise is inevitable for a wristwatch. When running/jogging, you usually have an angled forearm and keep the forerunner more or less in a horizontal position, so GPS sensor is in the presumably best possible position. But when walking, your arms just hang down, GPS sensor is in an unfavourable position. Detailed it out in this thread BTW.

Now fasten your belts, you don't see such a testequipment every day. :D
Image

regards
CHnuschti
 
Posts: 267
Joined: Tue Feb 13, 2007 3:28 am
Location: Switzerland

Postby pkan » Sun Aug 23, 2009 1:12 pm

My view of distances from 1s vs smart recording was based on tests I did about three years ago right after I bought my FR305. I basically looked at a lap that I often include in my runs, which means that I had a decent amount of statistics to look at, and compared 1s and smart recording. This lap is on a park road, half lightly tree-covered and half mostly open. At the time, it was rather clear that the ST distance numbers where inflated by about 2-3% with 1s recording compared to smart recording, while the GF305 numbers didn't differ between 1s and smart recording and were consistent with the ST smart recording numbers. Since then I've always used smart recording for outdoor activities.

After reading CHnuschti's post above I decided it was time to do some new tests and see if things have changed. To establish a baseline I took a closer look at data from the last few months and based on 29 laps with smart recording I got an average of 3357 m for the distances from ST and 3355 m for the distances from the FR305 (taken from the DistanceMarkers in my logbook), i.e., a difference of only 2 m (=0.06%). The standard deviation was 11 m (=0.3%) for both ST and FR305. Note that some of the variation may come from the runs not being quite identical. The standard deviation of the difference between ST and the FR305 was 6 m (=0.2%).

I've run 2x2 laps with 1s recording, which is not enough for a proper statistical analysis but can still give some indication. The distances from ST were 3346/3354/3378/3384 m and from FR305 3324/3338/3334/3351 m, which gives averages of 3366 m and 3337 m, i.e., a difference of 29 m (=0.9%). The FR305 distances seem a bit short compared to the smart recording data, but more data is needed to determine if this is indeed the case. Anyway, it seems clear that the difference between ST and FR305 is not consistent with what I found three years ago.

I've also tried walking the same lap once with 1s recording, and got 3439 m in ST and 3320 m from FR305, i.e., a difference of 119 m or 3.5%, with the ST distance being at least 2% more than I get for running. This seems statistically signficant even considering that I just have data from a single lap, but it is still far less than the 10% overestimation I had expected.

Looking more closely at the track data, it seems clear that there is a lot less false zig-zagging in the recent data than in the data from three years ago, and that whatever zig-zagging remains occurs at a slower time-scale, i.e., not from measurement to measurement. A quick look at old data also shows that reported distances from ST were roughly 1% larger overall, which is presumably due to the extra false zig-zagging.

So the distance accuracy in ST with 1s recording, and in fact the GPS accuracy of the FR305 in general, looks a lot better now than it did three years ago. This could be hardware-related, as my FR305 has been replaced, but could also be due to firmware updates. For instance, if I recall correctly the FR305 originally couldn't acquire satellites on the run, which meant that satellites could only be lost, never gained, once the run was started.
pkan
Donated!
Donated!
 
Posts: 145
Joined: Fri May 12, 2006 2:26 pm
Location: Sweden


Return to Questions

Who is online

Users browsing this forum: No registered users and 2 guests

cron