Why is 10 Hz GNSS (GPS) not enough for performance analysis?

Back to the overview
Benedikt, May 5 2026

Most dedicated sport performance trackers record an athlete's speed and position at, at best, 10 Hz. That is, 10 measurements (samples) per second. But is this enough? Can we really analyse an athlete's performance with only 10 samples per second?

No.

In this article I present four practical examples, comparing 10 Hz GNSS (GPS) data with data from our sensor fusion at 200 Hz (200 samples per second). I show what is measured at the sensor location on the upper back and how this differs from the motion of the centre of mass:

After reading this article, you will understand (1) what you miss if you only work with 10 samples per second and (2) the difference between the speed of the sensor (on the upper back, between the shoulder blades) and the speed of the centre of mass, obtained with a simple centre-of-mass body model.

GNSS or GPS?

GNSS stands for Global Navigation Satellite System. It's the overall term for all satellite-based navigation systems, such as GPS run by the United States, Galileo run by the European Union, BeiDou run by China, or GLONASS run by Russia. Our sensors always use the satellites from all available systems (the more satellites, the better. In good conditions our sensors receive over 30 simultaneously). This is why we always use the term GNSS and not GPS.

Centre-of-Mass Body Model?

From an energy point of view, you don't care how fast the sensor is moving or where in space it is. You want to know how the centre of mass is moving. And, in humans, the centre of mass is not on the upper back between the shoulder blades but, if you are standing upright, approximately at your belly button. It can shift depending on how you position your arms, legs, and bend your trunk. It can even be outside our body, the most famous example being the Fosbury flop in high jump.

Illustration of the centre-of-mass body model

It's this centre of mass you need to track to understand performance. The upper body may do additional movements, like sideways leaning or trunk flexion, adding considerable extra movement with respect to the centre of mass, and, therefore, masking what you actually want to measure.

Part of my PhD thesis was on this topic. If you are interested, you can access the main publication from 2016 "Three-Dimensional Body and Centre of Mass Kinematics in Alpine Ski Racing Using Differential GNSS and Inertial Sensors" here: doi.org/10.3390/rs8080671

How bad can this extra movement be? Very bad. Jump to the cross-country skiing section for an impressive example.

What is Sensor Fusion?

Sensor fusion is the process of intelligently combining data from multiple sensors so that each one compensates for the others' weaknesses and you get a more accurate, robust signal than from any single sensor alone or through interpolation. In our case, data from the GNSS, inertial sensor (accelerometer and gyroscope), and barometric pressure sensor is fused, resulting in drift-free speed and position at high time resolution, with GNSS measurement noise effectively filtered out and the advantages of all sensors preserved.

Athletics: 100 m Hurdles

It’s a numbers game (if you don't like numbers, jump this section)

As the "extreme case", consider 100 m hurdles: over a total distance of 100 m there are 10 hurdles, evenly spaced at 8.5 m (13m runup and 10.5m home stretch). The best women cover this distance in well under 12.5 seconds. Tobi Amusan (NGR) holds the world record at 12.12 seconds, from 2022 (Source). For the sake of simplicity, suppose the athlete runs at a constant speed of 8 m/s. Thus, the hurdle-to-hurdle time is 1.0625 seconds (8.5 m divided by 8 m/s). Between each hurdle the athlete does four steps. Ignoring that the one step over the hurdles lasts approximately twice as long, the average step duration is approximately 0.26 seconds. In other words, we have step frequency of 3.76 Hz (steps per second). This means that, when measuring at 10 Hz, we have a bit more than two samples per second. In a perfect world, according to the Nyquist-Shannon theorem, this is enough and, simplified, no more information is gained at a higher rate. However, is it truly enough considering that when we start looking more closely, during each step we have contact phase and a flight phase? Suddenly for each of these phases we have less than two samples. But shouldn't we not also have at least to samples during each of these phases? Whoops, we have a problem: the motion is much faster than the step frequency! Sampling at 10 Hz is too slow and cannot measure the whole movement!

An example from a short hurdles training session

First, let's start without the centre-of-mass model, looking just at how fast the sensor is moving. The figure below compares the ground speed measured at 10 Hz (i.e., 10 samples per second) with the GNSS (black line, each sample is marked with a small dot) with the 200 Hz speed from the sensor fusion (blue line). Both lines follow approximately the same pattern. Until about second 2.0 we have zero speed: the athlete is at the start and waits for the signal. Then we have the acceleration phase for about 2 seconds, followed by the steady state phase with the hurdles for the rest of the time.

Figure comparing the 10 Hz raw GNSS ground speed with the fused speed at 200 Hz

The GNSS ground-speed signal appears very noisy and shows regular spikes. These spikes are measurement artifacts caused by the abrupt changes in the athlete's upper-body inclination when jumping over each hurdle. For a very short moment, the sensor suddenly "sees" a different portion of the sky and receives a different combination of satellite signals. This makes the receiver briefly "think" it has moved much faster than it actually has, which shows up as a sharp speed spike in the data.

The 200 Hz speed signal looks much smoother and more regular. However, it still appears noisy, and you can only vaguely recognize the individual steps. This residual noise is mainly caused by the additional motion of the upper body, especially when clearing each hurdle (in this example, the average time from hurdle to hurdle was 1.15 seconds).

Now let's see what happens when we apply our centre-of-mass body model. The first figure below shows the raw 10 Hz GNSS ground speed in grey and the 200 Hz centre of mass ground speed in green. Because the difference may be hard to see, we also display the plot side by side to highlight it (second figure below). Suddenly, the structure in the 200 Hz data emerges: individual steps become easy to identify, and the steps over the hurdles stand out with a clear, repeatable pattern in the athlete's speed between the hurdles. Doing sensor fusion and applying a centre-of-mass model completely changes the signal: noise and measurement errors are removed, the individual steps appear.

Figure comparing the 10 Hz raw GNSS ground speed with center-of-mass speed at 200 Hz
Side by side comparison of the 10 Hz raw GNSS ground speed with center-of-mass speed at 200 Hz

You may ask why there are such large speed fluctuations of almost 5 km/h within each step? The main reason is air-drag: you can only accelerate when your foot is pushing on the ground. But when running / sprinting, during approximately 50% of the step, no foot is touching the ground. You are in the so-called flight phase. The air-drag slows you down and makes you lose speed. Air-drag increases with the square of the speed: the faster you go the more speed you lose. And, of course, the longer you take going over a hurdle the more speed you lose, too.

Cross-Country Skiing

Let's slow down a little bit and move to cross-country skiing. Usually, this movement is a bit slower. Thus, one could think that having only 10 Hz GNSS data would be sufficient to accurately measure speed. The figure below is from 30 seconds of cross-country skating. As for the hurdles example, the black line shows the 10 Hz GNSS speed and the blue line the 200 Hz speed obtained from sensor fusion. As cross-country skiing has considerable uphill and downhill parts we compute the 3D speed, also taking into account the vertical speed component.

Comparing the 10 Hz raw GNSS speed with the 200 Hz fused speed for cross-country skiing

Until approximately second 200, the athlete is going uphill. Then there is a transition, and at around second 208 a downhill section starts and the athlete goes in a tucked position. The curves seem to follow each other well for most of the time. The GNSS speed seems a bit noisier, with higher peaks and lower valleys. In the transition phase the GNSS speed has a bit of its own life.

However, remember that in cross country skiing the upper body has considerable extra movements and both the legs and arm are contributing propulsive power ("forwards force"). The figure below shows what happens when the centre-of-mass model is applied (green curve). For the last seven seconds displayed, when the athlete is in the tucked position, there is no difference: the upper body is stable and does not move. However, for the uphill and transition part where active skating takes place, the speed is suddenly completely different: the large fluctuations disappear. They are an artefact from the significant upper body movement and are unrelated to performance.

Comparing the 10 Hz raw GNSS speed with the 200 Hz fused speed and centre-of-mass speed for cross-country skiing
Comparing the 10 Hz raw GNSS speed with the 200 Hz centre-of-mass speed for cross-country skiing

Rowing

In rowing the sensor is attached to the boat, measuring how the boat moves through the water. One would think that it's a very smooth movement and that, surely, measuring speed ten times per second is more than enough? Well, no. Measuring rowing movement at 10 Hz is not enough.

There are two main reasons:
  • GNSS measurement noise
  • Movement frequency

We have observed that GNSS measurement noise is not strictly white noise. We can usually observe higher signal amplitudes, meaning that speed peaks are overestimate and speed minima underestimated. But randomly and now always the same way. Thus, if one would like to measure total speed changes, the GNSS will overestimate the true speed changes but we do not know by how much as it will be different for each stroke. The next figure, showing 16 seconds of rowing boat speed, shows this clearly: the raw GNSS speed (black line) sometimes overshoots the fused speed (blue line) but not always the same way.

Comparing the 10 Hz raw GNSS speed with the 200 Hz speed for rowing

We could try to low pass filter the raw GNSS speed, for example with a 2 Hz cut-off, as shown in the next figure. Although this smoothing reduces some of the visible noise, it also removes important performance information about how the boat moves through the water. At the same time, the problem of "signal overshooting" remains. In other words, low pass filtering does not bring the GNSS speed any closer to the speed obtained from sensor fusion.

Comparing the 10 Hz filtered GNSS speed with the 200 Hz speed for rowing

A movement is always faster than its cycle frequency (for example the stroke rate, cadence, or step frequency), because each cycle must be further divided into its different phases. In running, there are stance and flight phases; in rowing, drive, recovery, and the check phase. Each phase has its own characteristics, and speed changes differently within each of them. Moreover, timing is key: when exactly does the next phase start? Without sufficient time resolution, in other words, without enough samples per second, this cannot be determined accurately.

To illustrate how much information is hiding within 0.1 second intervals (the duration between two GNSS samples), we do a little experiment and compute the boat acceleration. To be fair with the GNSS data we have interpolated the 10 Hz GNSS speed with a cubic spline interpolator to 200 Hz prior to deriving the signal (five-point derivation) to obtain boat speed. To reduce interpolation artefacts the signal was further low-pass filtered at 4 Hz with a 2nd order Butterworth filter.

The next figure compares the results: the black curve is the GNSS-speed derived acceleration and the blue curve the true boat acceleration at 200 Hz. (Read our article on boat acceleration to understand this curve). Both curves have the same pattern. However, as expected, the GNSS curve has much more noise despite the applied low-pass filter. And key performance indicators are wrong or even completely lost. To name a few:

  • The acceleration minima at catch (the "canyons") are too low by about 20-30%.
  • The little acceleration dip just after the zero crossing after the catch, a key indicator for non-ideal rowing technique, is almost never visible in the GNSS-derived boat speed changes.
  • The end of the drive phase is almost never visible in the GNSS-derived boat speed changes.
  • The positive boat acceleration during the recovery is completely wrong in the GNSS-derived boat speed.
Comparing the boat acceleration obtained with the GNSS speed with the true boat acceleration sampled at 200 Hz

Snowboard-Cross

Snowboard-cross is probably the "smoothest" sport we have, even though speed is much higher than in athletics, cross-country skiing, or rowing. There are only few turns, and turns are usually very wide. However, there are jumps and other elements that require fast adaptation.

Is 10 Hz enough in snowboard-cross and related disciplines such as ski cross and alpine ski racing (slalom, giant slalom, super-G, downhill)? Only if you are happy with approximate and smoothed values and if you can coach with not seeing short-duration speed differences.

As for the other sports, we start again with a comparison between the 10 Hz raw GNSS speed (here again in 3D, also taking into account the vertical speed) shown in black and the 200 Hz fused speed in blue. There are considerable differences in the two curves. Why? It's the same as with the example we showed in athletics: GNSS measurement error because of changes in the upper body orientation which influences which satellites are visible to the sensor.

Comparing the 10 Hz raw GNSS speed with the 200 Hz speed for snowboard-cross

Especially the first part has large signal differences, in places over 5 km/h. That's a difference of almost 10%! The video below shows the athlete during the first few seconds of this plot: he is moving over bumps. Each bump creates a small "peak" in the signal, because of the extra movement of the upper body. In the GNSS speed it's sometimes here, sometimes exaggerated and sometimes absent. Therefore, completely unreliable if we want measure how an athlete moves over such elements.

Now let's see what happens when we remove the effect of this upper body motion by applying our centre-of-mass model. You will probably not be surprised anymore: the curve becomes smoother and the apparent "peaks" are reduced (though they do not disappear completely, because the bumps genuinely cause speed changes due to the changing terrain). Notice also how the speed before and after second 28 looks quite different: it is smoother and shows fewer fluctuations. This is because we now measure the centre of mass, which is much more stable than a point on the athlete's upper back. If we assume there is a jump just before second 28, the extra upper body movement alone can create speed differences of several km/h in the sensor signal. When comparing two athletes with different movement patterns, this can produce a non-existent speed difference at the jump of a few km/h and lead to completely wrong conclusions.

Comparing the 10 Hz raw GNSS speed with the 200 Hz fused and centre-of-mass speed for snowboard-cross

Conclusion

These examples show that 10 Hz GNSS alone is fundamentally insufficient for serious performance analysis. Movements in athletics, cross country skiing, rowing, and snowboard cross all contain fast phases and complex body motions that happen much quicker than one snapshot every 0.1 seconds. As a result, raw or even low pass filtered GNSS speed systematically hides or distorts key details such as step to step speed changes, phase timing, and the effect of technique on performance. By combining GNSS with inertial sensors and applying a centre-of-mass body model, we recover a much more accurate and stable picture of how the athlete truly moves. Once you have seen this level of detail, it becomes clear that high rate sensor fusion and centre of mass modelling are not "nice to have" extras, but essential tools for trustworthy performance analysis.