Around the year 2000, it became possible to track players’ positions in game sports during

match. This was a revolution in performance analysis (PA) because from then on, we had a
totally new family of data at our regular disposal. Before that, in the very beginnings of PA as
scientific discipline, it was only in principle possible to assess positions of players on the
pitch, for example, in projecting a motion picture or video frame by frame, mark positions
manually on a table, read the marks, and transfer the marks to a coordinate system (Winkler
1985). Another method used by PA pioneers was to watch players’ motions in a film or video
and trace it with a pencil on a football pitch drawn on a sheet of paper. Distance covered was
then assessed with a nowadays forgotten device that was called curve ruler. A further
approach was so-called time-motion analysis (Reilly and Thomas 1976). Methodologically,
these approaches have more in common with observational systems than with modern
tracking technologies. A human observer records the location (including distance) and type
of a player’s movements on the pitch. Although there was more and more support by
electronic devices, these time-motion analyses remained error-prone and very laborious,
because each player had to be “tracked” separately based on video footage (Bloomfield et
al. 2007). There was no way for a regular use of position data in practice, whereas today we
have GPS (Global Positioning System)-, radar-, and video-based position detection at work
in several professional game sports, in Europe dominantly in football and handball.

Performance Analysis and Technological Progress The example of position detection shows
very clearly that our capabilities in performance analysis depend on the state of the art of
available technology. Technological progress contains several sub-processes, such as
enhanced performance of technical devices, for example, Moore’s law on computing power
(Moore 1965), miniaturization, and cost reduction, all of which have led to the new
technologies becoming of value for performance analysis, finally. Technological innovations
are the result of a continuous process that is by no means driven by sports science. Instead,
car industry, entertainment industry, and of course military are the most powerful drivers.
Nevertheless, we may expect from technological progress a continuous stream of
innovations and experience shows that we may expect some of them showing potential for
performance analysis. It is the task of sports science to track technological progress, to
identify developments with potential for PA, to conduct pilot studies, and eventually to
introduce the innovation in practice on a routine base. National scientific support systems for
top-level sports may be evaluated by the degree they stimulate and promote activities aiming
to bring innovations to work in sports practice (Lames et al. 2016). For sports science as well
as for its discipline performance analysis this means that technological progress will keep us
continuously busy in the future and also that today’s methodological options, not so much
the concepts, presented in this book will be called outdated in some future.

This chapter starts with an introduction in the different technologies nowadays in use for
position detection in game sports. The focus here is only on technical details of raw data
acquisition and signal processing that are relevant to practical decisions in performance
analysis. Intricate details from an engineering perspective are left apart. Nevertheless, much
like in action detection, caring for reliability and validity of position detection is a relevant
issue for performance analysis as well and treated in a second section. The focus here is on
specific problems, for example, with appropriate gold standards or compatibility of position
data obtained from different systems/ technologies. Several examples for position-based
studies may be found in Chap.

3.1 Functioning of position tracking

3.1.1 Position Detection Methods
As mentioned above, at the current state of performance analysis, three major technologies
are used for position detection of players and/or ball: GPS-, radar-, and video-based tracking
(VBT). The presentation of these technologies will give a rough idea of their technical
functioning and then specify the requirements and limitations for each system for sports
analyses and finally line out typical use cases in sports.

3.1.1.1 GPS
GPS (Global Positioning System) is a satellite-based position tracking technology. Actually,
there are several systems with global position tracking capabilities (GLONASS, Russia;
Galileo, European Union; BEIDOU, China; Japan and India only regional) besides
NAVSTAR/GPS, which is the US system. Since the US GPS.

was the first one, it is open to each user and the most used system internationally; GPS has
become a synonym for satellite-based position tracking, although the correct label would be
GNSS (Global Navigation Satellite System).
The central feature of GPS is a network of orbiting satellites. A minimum of 25 satellites at
20,000 km height with inclined trajectories ensures that at least 8 satellites are “visible” at
most points on the earth’s surface (Hofmann-Wellenhof et al. 2012). The satellites are
equipped with an atomic clock and high-precision position measurement devices. They are
continuously emitting radio signals with a fixed and documented structure, containing roughly
spoken their ID, their actual position and a time stamp.
GPS sensors are able to receive these signals from “visible” satellites, that is, they require a
direct “line of sight” to them. By triangulation, each receiver is able to calculate its position on
the earth’s surface. This works in principle quite simple, because if the time a signal has
travelled from a certain position in space to a position on earth is known, one may calculate
the distance to this position and hence conclude that the receiver is located on a sphere
around the satellite with a known radius. If two of these spheres are known, one’s own
position is located on a circle formed by the intersecting spheres. Finally, the intersection of
the first two spheres with a third sphere mathematically gives the exact point of the receiver’s
location.
As one might assume already, in practice, things are not as simple. Electromagnetic radio
signals emitted by the satellites travel with speed of light (about 300,000 km/s), the satellites
themselves travel with very high speed, and gravitation has also a relativistic influence on
the measurements. Although mathematically the minimum requirement for a precise
triangulation in space is 3 satellites (see Fig. 3.1), the many sources of lacking precision
require a fourth satellite for appropriate time
estimation and an abundance of further visible satellites to improve accuracy. Typically at
least 8 visible satellites are required.
The precision of standard GPS is given with an upper 95% limit of 7.8 m with an RMSE of 4
m. Therefore, in sports, we typically use differential GPS that requires an additional receiver
with a fixed and known position typically close to the pitch. With the help of the signals of this
stationary receiver, one is able to clean the signal and get a better spatial resolution down to
some centimetres. A big improvement in the accuracy of speed data obtained by GPS came
when Doppler effect in the satellite signals, that is, an increase/decrease in wave frequency
when the sensor is approaching/receding from the source of the signal, was exploited for
assessing speed and direction of the moving sensor. At this point, it becomes obvious that
only with the advent of highly sensitive electronics and only in the recent years, one has
become able to conduct the sophisticated measurements associated with GPS.
First applications of GPS in sports came quite early for navigation in outdoor sports, such
as sailing, for example. For jogging and golf, there are successful GPS tools as well. In the
last years, we saw an increasing measurement frequency for GPS devices in game sports
starting from modest 1 Hz to 15 Hz and more (but in some devices only obtained by
interpolation from real measurements!) nowadays. This increase and the advent of
differential GPS and Doppler speed made GPS- based position detection in game sports
competitive to other technologies.
GPS performance is hindered by anything that prevents the receiver from getting the satellite
signals properly. Electromagnetic waves cannot penetrate metal or concrete structures.
Therefore, GPS position detection is not suited for indoor events. Although there are some
technical remedies under way, this remains a strong limitation for tracking sports events, not
only indoor events but also the very relevant outdoor events that take place in bowl-shaped
or even roofed stadia.
This is unaffected by the option of collecting data from GPS sensors equipped with Inertial
Measurement Units (IMUs) that provide indoor data, also. When using GPS outdoor, there
still might be problems caused by so-called multipath effects; that means a signal from a
satellite is reflected by rocks, buildings, metal constructions, or wet surfaces, and a receiver
gets two interfering signals reducing accuracy. Optimal conditions for GPS are given by a flat
plane with unobstructed view to the sky and no metal or concrete or wet objects around. One
may assume, for example, that GPS works better on a training pitch than in a football
stadium. In a recent study (Shergill et al. 2021), the quality of the GPS signal was investigated
over 50 matches in 24 stadia in English Championship league (second league). It found on
average sufficient conditions, for example, the number of satellites visible, their distribution
(HDOP = horizontal dilution of precision), but an average of 11 min of playing time was
measured regularily, where the recommended signal quality thresholds of the provider were
not met. Also, there were significant differences in results depending on the quality of the
GPS signal.
Moreover, a requirement for GPS-based position detection is a receiver attached to the
tracked object. In recent years, these receivers became very small and light and can be easily
worn in a pocket of a bra-like harness typically between shoulders (line of sight!). Problems
occur when the ball must be equipped with a GPS sensor or when only sub- optimal fixation
(hips, below protecting sports harness) is available. A big advantage of GPS for applications
in sports is its ready-to-use property. As soon as receivers are attached and the base station
is plugged in and records data from the receivers, the system provides data. Also, GPS is
comparatively cheap, although costs increase with increasing accuracy demands.

3.1.1.2 LPS
Local Positioning Systems (LPS) work similar to GPS. The decisive difference is that LPS do
not work with satellites but create their own reference system by distributing so-called base
stations around the pitch. These base stations receive signals from a sensor worn by the
tracked object. Unlike GPS sensors, LPS sensors receive and actively emit signals. They are
called transponders (transmitter/responders) or RFID (radio-frequency identification) chips.
Again, the position of the transponder is obtained by triangulation of time-of-flight information
from several base stations. Since a typical time of flight is now only some nanoseconds (1
ns = 0.000,000,001 s), several electronic measures must be taken to achieve sufficient
accuracy, including an additional stationary transponder and the reference transponder, with
a known position. These systems are also called “radar”or “radio”-based position detection
systems, because typically the frequencies of their signals are located somewhere between
radar and radio in the frequency spectrum of electromagnetic waves.
First applications of LPS in sports started in the first decade of this century (Stelzer et al.
2004). Initially, it was used for race car tracking and motorcycle tracking in moto-cross events.
Soon, LPS was applied in soccer, speed skating (Drawer 2008), and several other sports,
too.
The precision of LPS, at least the one given by the manufacturer, is usually quite high, for
example, an error rate of 5– 10 cm is reported (www.inmotio.eu). Another big advantage of
LPS compared to GPS is the sampling rate. Typically, we have a basic sampling rate for a
system (e.g. 1,000 Hz) that must be divided by the number of the sensors in the field.
Tracking 22 football players thus is possible with roughly (reference sensors, substitutes)
1,000 Hz/22 = 45 Hz.
Ball tracking with LPS requires the equipment of a ball with a sensor which remains quite a
challenge and has reported to be successful by only a few manufacturers with only some
documented reliability studies so far (Seidl et al. 2016; Blauberger et al. 2021). Nevertheless,
ball tracking via LPS is available for football, basketball, and handball and will enhance
diagnostic options in the near future.
The practical requirements for a LPS allow indoor position detection. There are considerable
space requirements because base stations need to be installed around the pitch with a
sufficient distance to the pitch; thus measurements in smaller indoor halls may create
problems. Moreover, the positions of base stations and reference transponders must be
known exactly. Both requirements, mounting and measuring of base stations give a
preference to fixed installations. Mobile measurements require the setup of base stations
and determining their position with a tachymeter with a precision of some millimetres. Since
the technology used is more demanding, costs for LPS are typically higher by a factor of 2–
3 compared to GPS.

3.1.1.3 Video
Video-based tracking (VBT) is a completely different technology compared to GPS and LPS.
It relies on video images of the match taken by one or more cameras. A video frame consists
of rows and columns, which are represented digitally by their colour value. For example, the
pioneering VGA standard had 640x480 pixels with 262,144 possible colours. More modern
standards are full HD (1.080 × 1.920 pixels) and 4 k or ultra HD (4.096 × 2.160 pixels). VBT
solves the task to determine the positions of objects based on colour information of
thousands of pixels per frame (VGA: 307,200 pixels; 4 k: 8,847,360 pixels) entering the
system at a frame rate of typically 25 Hz (PAL; NTSC: 30 Hz). Object identification is
addressed by making use of several image processing technologies that became only
available in the last decades due to progress in the capacities of relevant hardware and to
theoretical innovations in image processing. The basic process may be described— very
simplified— with four steps: video recording, world-to-pixel transformation, object
segmentation, and pixel-to-world transformation (Beetz et al. 2005).
The quality of video recordings is a decisive prerequisite for the quality of VBT. We need a
full-pitch coverage allowing the identification of the objects of interest. First, the tracked
objects must be represented with a sufficient number of pixels to make it possible for the
algorithms to detect the object. This might be achieved by using highresolution video (e.g. 4
k), but this used to be limited in the past by the amount of data to be processed in real time.
Although it is in principle possible to analyse videos from a swaying, tilting, and zooming
camera, typically several fixed cameras are used to cover the whole pitch. We know systems
with two cameras as well as systems with 20 or more cameras, the latter solutions again
touching technological limits of data handling. A second problem may at present only be
alleviated, not solved: occlusions, that is, one object (player) appears in front of a second
one in the video image. Thus, the cameras should be distributed around the pitch, and also
their mounting position should provide a reasonable angle of sight to the pitch so that the
number and degree of occlusions are minimized (see Fig. 3.2).
The next step consists of the world-to-pixel transformation f: world → pixel. This means that
for each camera, the location of each point on the pitch in the pixel matrix has to be
established by a calibration process. For this purpose, objects with known location are
identified in the pixel matrix. In football, typically the lines on the field are used with their
known (middle circle, penalty box, penalty point, goals) or assessed (length and width)
dimensions. Problems that may occur at this step are optical bias of the camera lenses and
a non-flat surface of the pitch, the latter being generally the case because for reasons of
water drainage, the elevation of the pitch may be up to 0.5 m higher in the middle compared
to the corners.
The central step of VBT is the identification of the objects of interest (players and ball) in the
pixel matrix. For this purpose, several techniques of pattern recognition are at hand (Beetz
et al. 2005). For example, background subtraction allows identifying so-called blobs, that is,
potential locations of players, by subtracting the current pixel matrix from an average value
per pixel obtained from several frames (background). The most important cues are the dress
colours of players of the two teams, which must also be “learned” before the analyses. A
player model consisting, for example, of head, shirt, trousers, and legs/socks with the
respective colour.

Fig. 3.2 Calculation of area with sufficient elevation angle (orange line) dependent on camera
position (blue). Parameters are elevation angle (here: 12°), camera’s distance to pitch (15
m), and height above pitch (20 m) values is fitted to each blob. With this player model, the
pixel with the best representation of the player’s position on the pitch (usually a vertical
projection of the player’s estimated body centre) is identified.
The world xy-coordinates of the player are obtained by making use of a pixel-to- world
transformation, which is the inverse of world-to-pixel transformation that was established
initially f − 1 : pixel → world.
It must be mentioned that these four steps are a very simplified description of the process.
In practice, information from previous frames and from the other camera views is necessary
to establish a good position estimate, and several more techniques of pattern recognition are
involved (Beetz et al. 2005).
As already mentioned, a problem in VBT is video quality. In a field setting, we find unfavorable
or changing light conditions with sharp and changing light/ shadow contrasts. Depending on
elevation angle, there is more or less background interference. Weather conditions such as
rain, fog, and snow should hamper VBT as well. Also, colours of player’s dresses may change
with ground contact especially on muddy playgrounds. The extent of influence on
measurement results of these effects is unknown since there is no study so far that has
investigated the impact of these problems.
The biggest and remaining problem for VBT is occlusion. This means that a blob contains
more than one relevant object. Although one tries to care for this as much as possible by
camera positions and blob analysis, occlusions are in some situations inevitable when
players approach each other very closely like, for example, in front of the goal during corners.
Even less relevant situations may lead to occlusion, for example, hugging players after a goal
scored. Intervention of human operators is needed to reassign the player’s identity to the
segmented objects. In contrast to the automatic GPS and LPS, VBT is only a semiautomatic
technology, requiring human intervention typically some hundred times per match in the first
years.
A big advantage of VBT is that it works without sensors carried by the objects. Up to 2015,
when IFAB and FIFA started to develop standards for electronic performance and tracking
systems (EPTS), any electronic sensors were prohibited in football, thus video-based
systems being the only alternative for recording positions in football. The same situation is
found in basketball where in NBA players were not allowed wearing LPS sensors until 2017
when wearable technology was in principle admitted.
The sampling frequency is identical to the video frame rate, that is, 25 Hz with PAL standard.
In principle, it would be possible to use high-frequency video, but as with high-resolution
video, this is limited by data processing capacities. Accuracy faces additional problems in
comparison to other technologies like distance to the cameras or direction of movement
relative to the camera axis (moving perpendicular or in line with the camera axis) (Siegle et
al. 2013).

The business model is different from GPS and LPS. While with GPS and LPS a customer
may purchase and run a system, VBT is typically offered as a service of specialized
companies. Typically such a company is contracted to supply a whole league or event with
position data as VBT is much more demanding what technological (data network) and
personal resources (correction of occlusions) is concerned.
Since there are high demands concerning the quality of video footage, the practical use of
video-based tracking is mostly limited to competitive matches in professional leagues in large
stadia. There is also a tradition of VBT-based indoor position tracking since around 2010 in
NBA. Only recently, systems enter the market that do tracking from broadcast video or use
drone-based videos thus becoming independent from fixed cameras at elevated positions.
In Table 3.1, a short comparison of the three tracking technologies is given. Each category
is roughly characterized with a four-level scale.

3.1.2 Signal Processing

Each of the three technologies mentioned above delivers as raw data continuous streams of
xy-positions of the relevant objects at their respective sampling frequency. This raw data is
typically still quite noisy and needs further processing to
obtain meaningful trajectories. This process is called smoothing or filtering. Most of the
tracking systems deliver only post processed, that is, smoothed data to the client. The client
does not come in touch with raw data. Ogris et al. give the rare opportunity to have a glance
at raw data, in this case from a LPS (Ogris et al. 2012, Fig. 3.3). It is very impressive to see
the great extent of raw data scattering and the heavy corrections imposed by smoothing, in
this case by a Kalman filter. Only after smoothing, the trajectory becomes comparable to the
gold standard trajectory at all.
Nevertheless, since post processing may be responsible for errors, especially in certain types
of movements, some aspects of it need to be mentioned here to create a good understanding
of the functioning of position detection in sports. In the following, the impact of sampling
frequency and degree of smoothing will be made aware of. Again, the focus is not on the
technical details but on consequences for position detection in performance analysis.

3.1.2.1 Sampling Frequency

The impact of the sampling frequency, that is, the Hz rate of the measurements (1 Hz is one
measurement per second), on the precision of measurements is best illustrated graphically.
In Fig. 3.3, the approximation of an original movement trajectory (red) by measurements of
different frequencies (blue, turquoise, green) is illustrated.
It is easy to perceive that the lower frequencies are not capable to pursue the original
movement satisfactorily and thus not capable to give accurate estimates, for

example, for the distance covered by a player on this trajectory. One might also arrive at the
conclusion that the necessary frequency depends on the original movement. If this is a less
curved line, like a jogging track, measurements with lower sampling rates would yield good
estimates, also. In signal theory, the minimum necessary sampling rate is given by the
Nyquist-Shannon sampling theorem (Shannon 1949). It says that this is twice the maximum
frequency of the frequency spectrum of a signal. This works only for signals given as a
mathematical function, because there is no upper limit for a spectral frequency in a natural
noisy signal.
There are some conclusions for the practice of position detection in sports, though.
Movement trajectories including sharp turns and high accelerations require higher sampling
frequencies than linear movements with lower accelerations. Thus, a low-frequency GPS
may be sufficient for monitoring the trajectory of a jogger. Competitive football, though, is—
at least concerning actions close to the ball— characterized exactly by the demanding
properties mentioned above, sharp turns and accelerations. This issue will be addressed
again in the next section on validation of position detection.

3.1.2.2 Smoothing
Since raw data from position detection are quite noisy (Ogris et al. 2012), they must be
mathematically processed before a reasonable trajectory of the object on the pitch is
obtained. If one would not do that, totally absurd values for distances travelled, speed and
especially acceleration would be reported. Smoothing is justified especially when taking into
account that each point of raw data is actually the most likely estimate for this point with other
estimates having a similar probability. There are several options available for smoothing, the
most popular ones are the following:

• Curve fitting: If we know that the movement is determined by a mathematical

function, we can determine the parameters of this function, for example, slope and
intercept of a linear function, by a “minimum least squares” (MLS) estimate.
• Moving average: For each point, the smoothed value is the mean value of some
points before and after that point. The number n of these points indicates the degree
of smoothing using moving averages, for example, MA (i , 25) gives as smoothed
value for point i the average of points i − 12, ..., i + 12.
• Spline regression: Splines are a series of cubic polynomials that interpolate a signal
with a curve of minimum curvature. In spline regression, a certain tolerance interval
around each point is allowed for obtaining an even smoother curve. The size of this
interval determines the degree of smoothing with splines.
• Filtering: This is a large family of smoothing methods that all have in common that
they work with the frequency spectrum of a signal. A frequency spectrum is the result
of a Fourier transform of a signal. A Fourier transform is a decomposition of a signal
into sine or cosine waves of different frequencies. Lower frequencies represent
components of the signal that change only over longer periods, whereas higher
frequencies represent rapid, rather discontinuous changes. In our case, trajectories
may be represented as frequency spectrum with lower frequen- cies standing for the
amounts of long-range movements such as basketball players moving between the
two baskets in the rhythm of the ball possessions of their own and the opposite team.
Higher frequencies in this spectrum are caused by rapid changes in the trajectory,
either “natural” ones, such as changes of direction and speed, or the changes from
measurement to measurement due to measurement error. With a high-pass filter, we
may eliminate all frequencies higher than a certain threshold from the signal and thus
get a smoothed (filtered) signal by applying the inverse Fourier transform. The
threshold for the high-pass filter determines the degree of smoothing.
As was pointed out, with each of the most common smoothing methods, there is in each case
a parameter determining the degree of smoothing. This degree is a compromise between
reducing measurement error (“noise”) as far as possible on the one hand and on the other
hand preserving the “real” trajectory. There is no general optimal solution for this problem.
Viable solutions depend on the amount of noise and on the characteristics of the real
trajectories. Especially if the latter ones contain abrupt changes in speed and direction, like
in many game sports, it is difficult to find a compromise.
Figure 3.4 illustrates this dilemma with a simulated example. The blue line is the original
movement, for example, the lateral displacement in one turn of a slalom run. The grey signal
is the blue line plus normally distributed random noise of a certain amplitude added; standing
for the raw data, one obtains as a concrete measurement. The orange line is the smoothed
grey signal (here a moving average over 25 measurement points was applied).
What we see is that smoothing gives (in this case) a fine approximation of the original
movement, but also, that in the surroundings of extreme values, here the change of direction,
smoothing leads to flattening the original peaks. This systematic problem tends to become
bigger with stronger smoothing and more extreme peaks in the original movement.
Summing up, one has to acknowledge that signal processing is an important part of position
tracking. Typically, the applied methods are hidden to the users of tracking data.
Nevertheless, it is useful to realize that there is no best way of signal processing, and the
optimal procedure depends on the error rate of the measurements and the properties of the
examined movements. Together with knowledge of the strengths and weaknesses of the
different tracking methods, this is a prerequisite for conducting adequate studies using
position tracking data.

3.2 Validation of tracking systems

Applied sciences like PA often face problems meeting academic standards. Especially the
validation of tracking systems may be criticized in this respect; details will be given explicitly
below. Typically, when new measurement devices are introduced in academics, validation
studies come first to demonstrate reliability and validity of the assessments. In sports
psychology, for example, one sometimes may have the impression that there are as many
papers on validating new questionnaires as papers with results obtained using these
questionnaires.
Performance analysis is dealing with positional data from professional sports competitions.
These are in general not collected by the researchers themselves but by commercial data
providers (see Box “Commercial Data Providers”). This frequently leads to the situation that
data without sufficient prior validation are available. Of course, it is seducing for a scientist—
including the author— to start analyses without a validation when data are at hand. Also, it
is understandable that there is not too much effort in ex post validation of widely introduced
procedures. Instead, methodologically poor “fig-leaf” validations are quoted, for example, Di
Salvo et al. (2006) with a validation of Prozone have 306 quotes according to Google Scholar
and Zubiaga (2006) with a validation of Amisco has 63 (a dissertation in Spanish!).
Nevertheless, an unsatisfactory academic level of validation studies is not only a problem for
the scientific discipline of performance analysis; it may lead also to validity problems with the
results of the studies and— very important for an applied discipline— it could lead to
inappropriate decisions being taken in the day-to- day work in practice.
Commercial Data Providers

When position tracking was introduced in the last two decades, this gave rise to a growing
industry of data providers. As this is a business, the products delivered are designed primarily
to meet the demands of the clients. Originally, the primary clients of data providers were
media, though. It is the media that pay for the information provided, in general via the owners
of the proprietary rights, such as clubs, leagues, or federations.

Since scientists do not act as direct customers of these companies, it is no surprise that
scientific standards do not play a big role for commercial providers:

- There is no interest in publishing the signal processing procedures involved, because

these are hidden from competitors as intellectual property.
- There is no interest in publishing the underlying definitions of the observational
systems, because this gives rise to never-ending discussions and justifications.
- There is no interest in objective, high-quality accuracy checks of position tracking,
because quality control may be achieved with internal testing as well and above all
without creating problems in case of bad results.

Taken together, there are good reasons why the companies in the past did not care
primarily for scientific standards, which in turn created methodological and ethical
problems for researchers. On the other hand, more and more clubs started to use
tracking information for practical purposes in training, thus relying on reliable data, and
the FIFA EPTS initiative heads in the same direction as well. At present, the accreditation
of EPTS by FIFA has become a routine, despite some open problems in EPTS validation
mentioned below.

At this point, it must be mentioned that practical demands require not only single systems
to be tested for accuracy, but also methods must be derived making positional data from
different source systems comparable. Why is this a relevant problem? Typically and
historically, match analyses are conducted by VBT. As explained above, the
requirements for VBT are frequently not given in training either for organizational reasons
(employ tracking companies) or for technical ones (camera positions). As a
consequence, position tracking in training and on training pitches is conducted using LPS
or GPS. If training pitches are equipped with stationary LPS, it is even possible that
sessions not accessible to LPS (different pitches or other locations, for example, training
before away matches) are tracked with GPS. In sum, we have at least two, sometimes
even more, tracking systems/technologies working for assessing movement profiles in
competition and training.
Moreover, as the purpose of training is preparation for competition, movement intensities
and other characteristics of exercises in training are frequently determined with respect
to match characteristics such as overall intensity and duration or number, intensity, and
duration of interval bouts (Buchheit et al. 2014) but also passing speed and number of
ball contacts. Taking these two aspects together— the use of different systems and the
need for cross-referencing between systems— gives rise to the demand for not only
validating individual EPTS but in addition also investigating the compatibility of different
position tracking systems.

Some naïve approaches (Buchheit et al. 2014) using linear transformations between the
data of two systems resulted in very unsatisfactory agreement between systems. It
became clear that deviations between systems are different for trajectories with different
characteristics, for example, linear jogging with low speed vs. sharp turns with maximal
acceleration. Linke and Lames (2019) found out that considerable differences between
systems originate in the different ways a player’s body is represented, for example,
sensor at the centre of shoulders (typical for GPS, mostly LPS) or the centre of pelvis
(sometimes LPS). VBT takes the centre of the detected player model which corresponds
to the centre of the pelvis. Taken together, comparability studies between EPTS pose a
specific methodological problem of their own.
This chapter continues with a decisive aspect of position tracking validation, the
employment of appropriate gold standards. As validation studies make up a special kind
of scientific studies, the next section is dedicated to questions of designing validation
studies in the sense of a general methodology. Finally, the present state of knowledge of
the accuracy of position tracking for different systems and different items is depicted.

3.2.1 Gold Standards for Position Tracking in Sports

The International Organization for Standardization (ISO; www.iso.org) defines a criterion
standard or gold standard as a value or measuring method that serves as an agreed-
upon reference for comparison, which is derived as a theoretical or established value
based on scientific principles. A gold standard is an indispensable ingredient for a
validation study, because something like ground truth is needed to compare the results
of the tested systems with (Luteberget and Gilgien 2020). The most appropriate gold
standard does not only depend on technological options but also on the kind of validity
that is demanded. Should the measurement be valid for whole body movements or just
centre of mass, are we interested in elementary movements such as linear runs or in real
football movements, and do we need validation of performance indicators or of each xy-
position? Taken together, it still is to be considered an art to find an appropriate gold
standard for a validation of a tracking system.
Despite these remarks underlining the necessity of gold standards in validation studies,
there are also studies that primarily compare results from different systems and do not
use a gold standard (e.g. Randers et al. (2010) and Harley et al. (2011)). In these cases,
one is not interested in the absolute accuracy but in a relative comparison of different
tracking systems. It might be a little bit unsatisfactory in cases of deviations between the
systems not to be able to tell which system is responsible for the (larger part of) errors,
especially one is not able to recommend a superior system.
As an important result, both studies demonstrate impressively that there are problems
with comparisons between different systems (although this becomes not obvious in the
discussion sections). Randers et al. (2010) examine four different systems and give
results that allow calculating the percentage of the lowest mean compared to the highest
mean (set to 100%) obtained from the four systems. In
detail, this is 87.8% for total distance, 60.8% for high intensity running, and only 54.8%
for sprinting distance. In a further paper on the topic, Buchheit et al. (2014) try to solve
the problem of incompatibility of data from EPTS employing three different technologies
by introducing linear transformations to adapt measurements from one system to the
others. Because of lacking agreement between the systems, this has to be done
separately for both different performance indicators and different pitch sizes.
The literature on validation of position tracking in sports knows several different gold
standards that were used in the past (Luteberget and Gilgien 2020):
- A very commonly used gold standard is a predefined movement circuit with known
spatial extensions (Frencken et al. 2010), for example, 5 m straight, 90° turn, and 5
m straight. Here, we may compare the “real” dimensions of the distance to cover (in
the example: 10 m) with the distance provided by the tracking system. The problem
with this kind of gold standard is that we are not sure whether the players actually
 perform the prescribed path exactly. In the case quoted above and in several similar
ones, there remain doubts, because the trajectory is— mathematically spoken— not
continuously differentiable, and it is physically not possible to run this path exactly at
constant speed (infinite angular acceleration required in turning point!).

- Timing gates are widely accepted as a high precision criterion measure; they provide
precise measurements of the time needed to travel from gate 1 to gate 2. As speed
reference, they are only of limited use (Aughey 2011; Redwood-Brown et al. 2012),
because one must again assume that the athlete actually performs the prescribed
path (which is a problem with nonlinear paths, for example, not continuously
differentiable paths with “corners”) and one must be aware that timing gates only
provide estimates of the average speed between the two gates. If one is interested
in the precision of xy-measurements, which is a quite natural question in testing
tracking devices, timing gates fail to answer it.

- Momentary xy-positions may be assessed using radaror laser-based distance

measurement devices . These devices, e.g. Laveg, emit radar or laser beams and
receive the reflection of a beam from the object very much like police does with cars
in traffic control. At a closer look, this gives only the distance of the object, but as this
is done at about 100 Hz, this provides a good estimate of momentary position and
speed as well. In tracking validation studies, this gold standard was applied, for
example, by Varley et al. (2012) and Siegle et al. (2013). The restriction here is on
linear movements. There may be acceleration or deceleration, but as only distance
is actually assessed, valid measurements are only obtained for movements directly
towards or away from the camera.

- Finally, the most advanced gold standard at present is “imported” from laboratory
research. Marker-based systems for high-precision movement assessment are
standard in lab-based movement analysis such as gait analysis. Technically, most of
them are working with infrared cameras and infrared-reflecting markers attached to
the object. In a lab setting, eight or more fixed cameras are installed, and a typical
measurement volume in a movement science lab is about 5 × 5 × 3 m. As this
technology is meant primarily for indoor use, transferring it to outdoor settings is quite
hazardous, for example, because sunlight “blinds” the infrared sensors of the
cameras and rain or even humid grass may give rise to reflections that may distort
marker detection.
At present, there are only few studies known using an infrared movement analysis system
as gold standard under outdoor conditions. Duffield et al. (2010) and Vickery et al. (2014)
validated GPS using exercises from tennis, cricket, and agility training, while Ogris et al.
(2012) and Stevens et al. (2014) tested the accuracy of LPS both using soccer-specific
movements. Linke et al. (2018a, 2020) used soccer- specific exercises also and tested GPS,
LPS, and VBT systems. These studies are characterized by using different numbers of
infrared cameras and different measurement volumes. The problem with infrared movement
analysis systems such as Vicon as gold standard for EPTS is the confined space of the
measurement volume. Duffield et al. (2010) use 22 cameras for (half?) a tennis court, Vickery
et al. (2014) use the same system for exercises with 21 m maximum length, Ogris et al.
(2012) use 8 cameras for a 26.5 × 24 m rectangle on a soccer pitch, and Stevens et al. (2014)
had a 10-camera system controlling exercises in a 30 × 3 m and a 15 × 6 m rectangle,
whereas Linke et al. (2018a, 2020) used 33 cameras for a 30 × 30 m square. A similar method
(25 x 25 m?) was adopted for FIFA’s EPTS validation.
A question that is frequently not addressed is controlling the accuracy of gold standards
(Merriaux et al. 2017). Frequently, it is not sufficient to report the manufacturer’s specification
instead of conducting measurements by oneself. Especially marker-based tracking systems
typically need to be installed in a field setting prior to each study. This requires sophisticated
calibration procedures, and thus even gold standard results are prone to errors. As a
consequence, the accuracy of the gold standard over the whole measurement area should
be demonstrated with a special investigation (Linke et al. 2018a, 2020).
A persisting problem up to now concerns the validation of EPTS in football. One would like
to be able to control position measurements all over a whole football pitch (FIFA norm: 105
× 68 m) with players performing a football match. The desire for a full-pitch gold standard is
not an aesthetic one, but we see maximum speed and acceleration values in the limited
measurement volumes that are much lower than in real matches. To make it even worse,
these high-intensity values are most error- prone for tracking and provide the most important
information for sports practice. Regrettably, one must state that in the studies published so
far, there is no gold standard allowing for full-pitch validation.

3.2.2 Design of Validation Studies

Conducting an EPTS validation study requires several methodological decisions, which can
be made more or less appropriately. Some of these are mentioned here to facilitate a critical
reception of validation studies. These decisions depend on the intention or purpose of the
validation (e.g. system optimization, system comparison,

demonstrating benchmarks) and thus are by no means trivial. The main sources for the
following sections are Linke et al. (2018a, 2020). As actual validation studies are conducted
almost exclusively in football, adaptation of the validation methodology to other sports is not
discussed.

3.2.2.1 Measurement Site

The measurement site should provide optimal conditions for the tested systems as well as
for the gold standard. A second criterion is ecological validity; the systems should be
scrutinized under conditions similar to their practical application. These claims together
create problems when EPTS using different technologies— GPS, LPS, and VBT— are to be
validated at the same time. The fact that each system has its specific optimal site (indoor,
outdoor, stadium, see Table 3.1) leads to compromises required when all systems are to be
tested at one time and— consequential— in one place. Buchheit et al. (2014) compared three
technologies using an open roof stadium and report non-optimal conditions for VBT. Linke et
al. (2018a) conducted their study in a wide stadium with only few spectators’ stands. They
erected an additional scaffold for VBT ensuring the required conditions (here: viewing angle).
The second study (2020) was performed in a professional soccer stadium (ESPRIT Arena,
Düsseldorf, Germany) for testing two VBT systems. The stadium’s roof was closed to allow
daytime testing with the infrared cameras of the gold standard system. Moreover, since
studies have shown an influence of the location on the pitch on the precision of the
measurements (Siegle et al. 2013), one either could choose a more difficult or more easy
location in the pitch for the tested devices. For VBT, a difficult condition is a location with a
maximum distance to the nearest camera and exercises conducted in a movement plane in
direction of the relevant camera’s line of sight. LPS might be hampered by concrete or metal
constructions beneath the pitch, whereas GPS could be affected by occluded line of sight to
the satellites, for example, by the tribunes of a stadium. As mentioned above, the choice
depends on the intention of the study; easy conditions show the full potential of a system,
whereas difficult conditions demonstrate its capability to scope with unfavourable conditions
that might occur in field settings.
3.2.2.2 Exercises
There is a broad range of exercises that could be chosen for EPTS validation. Typically one
will work with a spectrum of exercises each selected for assessing specific aspects of position
tracking. These aspects could be to examine elementary movements vs. football movements,
aiming in the first case to pin down performance of the EPTS to controlled elementary
conditions such as standing still, moving with continuous speed at different levels,
accelerations and decelerations, or continuous or accelerated curved runs. On the other end
of the spectrum, there are typical training exercises or (small sided) football games. Another
idea is to use exercises to be assessed more easy or more difficult by the EPTS. For
example, one might provoke occlusions to give VBT a hard job (Siegle et al. 2013). Above
all, as there is the limitation that we do not know a gold standard covering a full football pitch
yet, it is impossible at present to employ the most relevant exercise, an 11 vs.

11 full-pitch football match. This is regrettable not only for reasons of ecological validity but
also because of the fact that high-intensity running is reduced on a limited pitch
(Casamichana et al. 2012). High-intensity exercises in turn are most error- prone for position
detection (see Table 3.2 below) and, above all, most relevant to physiological assessments!

3.2.2.3 Levels of Analysis

The most basic information that is provided by EPTS is the xy-position of the tracked object.
This information is given at a certain measurement frequency specific to the system.
Obviously, it is a good idea to check for the precision of this basic information. On the other
hand, there are other variables derived from this information at an intermediate level such as
speed and acceleration, which are typically derived from xy-positions with the exception of
GPS using Doppler shift for direct speed assessment. As these data require additional data
processing and are subject to different measurement problems, it is recommended to
account for this intermediate level in a validation, too. Finally, performance indicators (PIs)
are derived from tracking data such as distance covered or number, distance covered, and
duration of stays in certain intensity intervals. Checking the validity of PIs performance
analysts in practice are working with then is also a required level of analysis.

3.2.2.4 Representation of Objects/Players

Position tracking systems allege to assess the position of an object, mostly the players on
the field. This position is given as a point estimate (xy-coordinate), whereas in reality, a player
is a three-dimensional object, of course. The question arises then, which xy-point is chosen
to represent a player position.

GPS and LPS locate the position of the player with the position of the sensors typically
attached to the centre of shoulders. VBT detects a model shape of the player and projects
the centroid of this model to the pitch as xy-position. Conducting a validation study, it is a
methodological consideration which point of the player captured by the gold standard is
actually compared to the EPTS. Each gold standard provides different points for comparison,
respectively different options. Whereas light barriers are shut by the first body part moving
through it, laser distance measurements track one point of the surface of the body (ideally;
with a proficient operator only!). Studies using Vicon typically put a marker on the sensor,
thus validating the sensor position. Linke et al. (2018a, 2020) attached five markers in their
studies allowing to assess not only the centre of shoulders but also the centre of the pelvis,
which is most of the time close to the centre of mass. In a special paper dedicated to the
representation problem, Linke and Lames (2019) found significant differences in the results
from both locations. This means, for example, that there will be differences between the
results of EPTS that are due merely to the fixation points of sensors. As a result for EPTS
validation methodology, we have two alternatives to test a system vs. a gold standard: The
gold standard tracks the system’s sensor positions, which is the fairest procedure from the
system’s perspective, or the gold standard tracks the centre of mass of the players and
compares this to the results of the tested system, which puts the accent on ecological validity.

3.2.2.5 Data Processing Steps

Last but not least, validation of EPTS is demanding with respect to the data processing
required to perform this task. Initially, we have xy-data streams from the systems under
scrutiny and the gold standard.
- Re-sampling: As systems have their individual measurement frequencies, the data
streams must be re-sampled (either up-sampled or down-sampled) to a common or
pairwise common sampling rate.
- Space synchronization: As each system has a specific origin and coordinate
system (e.g. GPS comes with the geographic coordinates of latitude and longitude),
spatial positions must be synchronized. This is done with a set of linear
transformations (e.g. rotations and translations), for example, with a generalized
Procrustes analysis (GPA; Gower and Dijksterhuis 2004).
- Time synchronization: The measurements must start at the same point in time to
be comparable. This may be achieved by event triggering, by electronic
synchronization of the devices, or by cross correlation synchronizing of the data using
the best fitting time lag.
- Smoothing: Typically, the xy-positions the systems provide have undergone an
internal smoothing, already. At the level of xy-data, it is a good idea just to compare
the output of the systems. At higher levels though, results such as speed and
acceleration may be obtained only after smoothing the positional data again. Since
this leads to differences in the results, for example, heavily smoothed trajectories are
shorter than mildly smoothed ones, a methodological consideration here is to treat
each data set with the same smoothing procedure. This provides a higher internal
validity to the comparisons, but practice may sometimes expect a comparison of the
net results of a tracking system stressing ecological validity.
- Performance indicators: Sometimes the systems make use of different definitions
of performance indicators. This may especially be the case with those that employ
speed thresholds for intensity intervals. These have to be homologized, of course,
prior to be able to conduct comparisons at the level of performance indicators.
After having made the data comparable with the above steps, a statistics assessing the
differences between the results of two tracking systems is needed. When one system is
accepted as gold standard, we may call this difference the error of a

tracking system. In the literature, we find several statistics used to characterize errors. The
classics in the field (Di Salvo et al. 2006; Zubiaga 2006) used correlation, for example. This
may be done only with caution; see Box “Correlation as Measure on Agreement”.
Correlation as Measure of Agreement

“Correlation is an instrument of the devil!” (Hilgard 1955, p. 228). Although this statement
was given with respect to a certain problem in personality psychology, the pitfalls of
correlation are in place until today as, for example, Aggarwal and Ranganathan (2016)
mention! Early classics in EPTS validation correlated velocity data obtained from timing gates
and the tested device. Because they merged results for different velocities, on one hand,
they obtained excellent correlations; on the other hand, they fell into the pitfall known as
correlations with heterogeneous subgroups. Simply spoken, the different running velocities
(differences over 10 m/s) make the errors appear small (large per cent errors, but <1 m/s),
thus leading to very high correlations. If correlations would have been reported per velocity
level, the reported scattergrams let expect very poor correlations.
Percent agreement, that is, how many per cent of the gold standard’s result does the tested
system provide, has the problem that error compensation is not prevented, that is, when
some parts of the measurement overestimate the true values and others underestimate
them, the per cent deviation does not give a clear picture. A point-to-point comparison is
achieved by the root mean square error (RMSE)

n, number of measurements; i , i th measurement; gs, gold standard; and syst, system tested.

RMSE gives an absolute value with the dimension of the measurement, for example, metres
for positions or metres per second when speed measurements are compared. By squaring
the differences, compensation is avoided and larger deviations are weighed higher. RMSE
may be interpreted as a something like a mean error (mathematically not true!), which is quite
intuitive.
In some cases, it is illustrative to give the RMSE as a percentage of the mean value of the
measurements (RMSE%). This provides a value that is comparable to a coefficient of
variation (CV) and may be taken as relative error. In Table 3.2, results from Linke et al.
(2018a) from the validation of a 2 min bout of a small sided game in football from a LPS are
given with RMSE statistics.
Table 3.2 shows that RMSE and RMSE% describe different aspects of accuracy. Whereas
RMSE, the absolute error, stays rather stable, RMSE% increases sharply with speed most
probably due to smaller amounts of distance covered in higher- intensity intervals. This
makes it very problematic to differentiate between player performances in these categories,
although the absolute error is not the biggest problem. Taken together, for the specific needs
of validation with a gold standard, RMSE- based error measures may be recommended.

3.2.3 Accuracy of position tracking in sports

This section is meant to give a summary of actual findings on accuracy of EPTS per
technology, thus including again some facts that were already mentioned in the technical and
design sections above. Doing so, one must be aware that things are under way and
companies permanently try to improve their products (Linke et al. 2020; Blauberger et al.
2021). Especially the integration of IMUs (inertial measurement units) into the sensors of
GPS and LPS will lead to substantial improvements for these technologies in the future.
Accelerometers and gyroscopes together with positional data will result in much better
position estimates than position detection technologies alone, because position
measurements may be combined with knowledge on the movement from the IMU sensors.
For example, one might expect that problems with high accelerations that all systems are still
facing will be alleviated when acceleration is known from the IMU and can be literally taken
into account.
Nevertheless, it is informative to collect the actual knowledge on error rates per technology
depending on exercise, performance indicator, and conditions of measurements. The
following results are taken from Linke et al. (2018a, 2020) if not mentioned otherwise.

3.2.3.1 Video-Based Tracking (VBT)

The accuracy of VBT depends on the position in the field. Siegle et al. (2013) showed that in
areas of the pitch being close to the cameras, results are much better than in areas far away
from cameras. A superimposed effect is movement direction, though. When movements are
in the line of sight of the cameras, the values mentioned above that were obtained from runs
perpendicular to the line of sight deteriorate significantly. It is important to note, though, that
the tested VBT system had all cameras installed at one spot. Linke et al. (2020) compare two
systems with cameras on one spot and distributed. They find improvements on data
completeness (fewer occlusions) but comparable accuracy.
A specific problem of VBT is the potential loss of identity of a tracked object in the case of
overlays. Supervision of human observers is needed resulting in only half-automatic position
tracking. One test condition employed by Siegle et al. (2011) pursued the strategy of
provoking identity problems: two runners sprint up to a meeting point and then stop, turn, and
sprint back to their starting position. The VBT under scrutiny reacted with considerable
increased errors. Of course, these were very unfavourable conditions, and maybe other VBT
systems employ better algorithms to compensate for temporal identity losses, but
nevertheless, VBT faces these problems that other technologies do not share.
Under favorable conditions, VBT shows errors around 20 centimetres for linear runs and 50
centimeters for football specific movements. The average errors in

speed measurement for football specific movements are 0.41 m/s; acceleration is assessed
by VBT with a typical error of 0.91 m/s 2 . In comparison to the other systems, these are the
biggest errors.

3.2.3.2 LPS
LPS require an around-the-pitch installation of base stations. Occasionally, different quality
of the measurements depending on position on the pitch is reported, but Siegle et al. (2013)
and Linke et al. (2018a) found LPS being remarkably unaffected by location on the pitch,
though.
Under favourable conditions, positional errors of LPS will stay below 20 cm for linear
movements and below 30 cm for football specific ones. LPS seems to be specifically
vulnerable in a run-stop-turn condition (RMSE 0.48– 0.60 m). The average errors in speed
measurement for football specific movements are 0.25 m/s; acceleration is assessed by VBT
with a typical error of 0.68 m/s 2 . In comparison to the other systems, these errors are among
the smallest ones.

Overall, LPS does not perform very well in measuring accelerations. There are two reasons
for that. First, acceleration must be calculated by twofold derivation of positional data and
each derivation is an error-prone operation. Second, position detection works with estimates
of the next position from positions in the past. These estimates perform great when the next
position may be well approximated by extrapolation of the former positions, that is, when
there was not much change in the trajectory, but they perform worse when this is not the
case, that is, big directional changes or accelerations/decelerations occur.
3.2.3.3 GPS
Accuracy of GPS is determined by the number of “visible” satellites. This requirement may
prohibit GPS measurements at all as is the case in indoor facilities. It may also cause
problems on certain areas on the pitch when, for example, walls or roofs interfere. In modern
football stadiums with high and steep grandstands, maybe even with a closable roof, large
problems have to be expected using GPS (Shergill et al. 2021).
In earlier times, low measurement frequencies of 1 Hz (later 5 Hz) were a problem with GPS
(e.g. Aughey (2011)). In the meantime, not only measurement frequency has improved but it
became possible also to use the Doppler effect for speed measurements.
As a result, we find GPS being very competitive concerning speed and acceleration
assessment with an error of 0.28 m/s for speed and 0.67 m/s 2 for acceleration
measurements in football specific exercises. Position detection is a weakness of GPS with
an RMSE of 0.96 m for football specific movements.
In Table 3.3, these results are summed up. Of course, one has to admit that only single
representatives of the three technologies were analysed; other systems may perform
differently. Also, it must be considered that comparability is limited because the figures are
sampled from two studies. The table thus only serves to illustrate the general trends in
accuracy performance within and between the different technologies.

Taken together, this section on accuracy of EPTS technologies shows that there are several
variables in the field exerting influence. Also, results depend on exercises employed and on
whether considering position, speed, or acceleration data. Consequently, different
performance indicators are affected differently by the technologies. For practical purposes,
this means to select very carefully among the options according to the local needs. At
present, the biggest problems for EPTS validation are the lack of a full-pitch gold standard
and the persisting problems with EPTS comparability.
In Chap. 6, it will be discussed which persisting problems and which future options will most
likely appear in the future of position detection. For example, a number of additional sensors
that may be used during matches will extend greatly the potential of PA, and maybe position
detection will be only perceived as the starting point of a new era of match analysis.