I have raw GPS locations for a track, and need to get maximum speed. Simple methods (calculate speed between 2 points, take max) will give meaningless large numbers due to GPS inaccuracy, the point is jumping around. Can you suggest a good ready-made algorithm to solve this?

  • If the point is jumping around are you saying you have more than one point at a location? The bulk of your question sounds as though you are interested in displaying real time data and not postprocessing (for accuracy).
    – Brad Nesom
    Commented Mar 16, 2011 at 21:44
  • If you have the 'raw' gps locations, then you should, by all means, have the speed over ground in knots which is a standard element in the NMEA RMC string.
    – nagytech
    Commented Jul 13, 2012 at 5:34

3 Answers 3


To get speed you must have time, of course. Thus you can order your points by time in a spreadsheet like fashion, with columns {Time, X, Y}, by increasing time.

Here is an example where the GPS unit almost completed a counterclockwise circuit:

Map of a trip

These points were not obtained at equal intervals of time. Therefore it is impossible from the map alone to estimate speeds. (To help you visualize this trip, though, I made sure to collect the gps values at almost equal intervals, so you can see that the trip started out fast and slowed at two intermediate points and at the end.)

Because you're interested in speed, compute the distances between successive rows as well as the time differences. Dividing distances by time differences gives instantaneous speed estimates. That's all there is to it. Let's look at a plot of those estimates versus time:

Plot of speed vs. time

The red points plot the speeds while the gray curve is a crude smooth, solely to guide the eye. The time of the maximum speed, and the maximum speed itself, are clear from the plot and readily obtained from the data so far if you're using a spreadsheet or simple data summary functions in a GIS. However, these speed estimates are suspect because the gps points clearly have some measurement error in them.

One way to cope with measurement error is to accumulate the distances between multiple time periods and use those to estimate times. For example, if the {Time difference, Distance} data previously computed are

d(Time) Distance
0.90        0.17
0.90        0.53
1.00        0.45
1.10        0.29
0.80        0.11

then the elapsed times and the total distances over two time periods are obtained by adding each pair of successive rows:

d(Time) Distance
1.80        0.70
1.90        0.98
2.10        0.74
1.90        0.40

Recompute the speeds for the accumulated times and distances.

One can carry out this calculation for any number of time periods, achieving ever smoother and more reliable plots at the cost of averaging out the speed estimates over longer periods of time. Here are plots of the same data computed for 3 and 5 time periods, respectively:

Plot of speed vs. time, 3 interval calculation

Plot of speed vs. time, 5 interval calculation

Notice how the maximum speed decreases with the amount of smoothing. This will always happen. There is no unique correct answer: how much you smooth depends on the variability in the measurements and on what time periods you want to estimate speeds. In this example you could report a maximum speed as high as 2.5 (based on successive GPS points), but it would be somewhat unreliable due to the errors in the GPS locations. You could report a maximum speed as low as 2.1 based on the five-period smooth.

This is a simple method but not necessarily the best. If we decompose GPS locational error into a component along the path and another component perpendicular to the path, we see that the components along the path do not affect the estimates of total distance traversed (provided the path is sufficiently well sampled: that is, you don't "cut corners"). The components perpendicular to the path increase the apparent distances. This potentially biases the estimate upward. However, when the typical distance between GPS readings is large compared to the typical distance error, the bias is small and is probably compensated for the tiny wiggles in the path that aren't captured by the GPS sequence (that is, some corner cutting is always done). Therefore it's probably not worthwhile developing a more sophisticated estimator to cope with these inherent biases, unless the GPS sampling frequency is very low compared to the frequency with which the path "wiggles" or the GPS measurement error is large.

For the record, we can show the true, correct result, because these are simulated data:

Plot of true speed versus time

Comparing this to the previous plots shows that in this particular case the maximum of the raw speeds overestimated the true maximum while the maximum of the five-period speeds was too low.

In general, when the GPS points are collected with high frequency, the maximum raw speed will likely be too high: it tends to overestimate the true maximum. To say more than this in any practical instance would require a fuller statistical analysis of the nature and size of the GPS errors, of the GPS collection frequency, and of the tortuousness of the underlying path.


This is not a ready-made script or algorithm. What I did in the image below showing average speed (in kph):

  1. Run a gpsbabel filter directly on the GPX file.
  2. Convert the GPX file into raster points in GRASS.
  3. Run r.neighbors to get the average speed for a specified raster window.

my average cycling speed

  • Could you clarify how it's possible to get speed from a raster representation of a sequence of locations?
    – whuber
    Commented Mar 16, 2011 at 14:42
  • You can synthesize speed information from the trackpoints of a GPX using gpsbabel. gpsbabel.org/htmldoc-development/filter_track.html. Then import this data as vector points and convert to raster in GRASS. I used multiple GPX on this image. Using r.neighbors, I get the average speed.
    – maning
    Commented Mar 17, 2011 at 1:51
  • Many thanks for the explanation. But, if gpsbabel has computed the speed, why do you use r.neighbors? Wouldn't that potentially mix speeds along one route with speeds along any other routes that are sufficiently close on the grid? Also, speed averaging is biased when GPS collection times are unevenly spaced. For example, if you travel 60 meters in 60 seconds, you have gone one meter per second, but if it's broken into 10 meters in 1 second plus 50 meters in 59 seconds (due to a +9 m positional error at a midpoint), the average speed computes to 5.4 m/s: a gross overestimate.
    – whuber
    Commented Mar 17, 2011 at 2:02
  • @whuber, you are correct. For this map, I'm not looking for individual speed but an aggregate speed through time. All my tracks are in 1 sec interval.
    – maning
    Commented Mar 17, 2011 at 2:17
  • +1 for a creative solution. (I like finding raster solutions to apparently vector problems, but it's good to know their limitations.)
    – whuber
    Commented Mar 17, 2011 at 2:52

Since your GPS data is inaccurate, you will only be able to estimate the maximum speed.

You can try to calculate it by computing the speed not on segments but on polylines (average speed), to minimize the effect of inaccuracies.

Have you tried cleaning your data first (Douglas-Peucker for example) to keep only the most relevant points ?

  • 1
    Douglas-Peucker does not really clean data - it just removes points while trying keeping original shape, so in fact it could even make errors look worse.
    – JaakL
    Commented Mar 16, 2011 at 13:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.