Using Individual Ball Possession as a Performance
Indicator in Soccer
Daniel Link
Hendrik Weber
Department of Performance Analysis
Technische Universitaet Muenchen
Muenchen, Germany
Department of Competition Information Data
Deutsche Fussball Liga GmbH
Frankfurt, Germany
[email protected]
[email protected]
ABSTRACT
This paper describes a model for detecting individual ball
possession in soccer based on positional data. The approach
used is able to determine how long the ball spends in the
sphere of influence of a player based on the distance between the players and the ball together with their direction
of motion, speed and the acceleration of the ball. When applied to the error-corrected raw data, the algorithm showed
an accuracy between 80 and 92%. As a first application
individual ball possession is used to calculate the invasions
to the opponents pitch side. This quantity is used as an
indicator for attacking and defending performance of teams.
Categories and Subject Descriptors
H.2 [Database Management]: Database Applications Data Mining
Keywords
Soccer, Performance Analysis, Performance Indicator, Individual Ball Possession, Invasion Index
1. INTRODUCTION
Performance analysis plays an important role in soccer
coaching. Observing and analysing tactical behaviour can
generate useful information that can be used for managing
training processes and developing match strategies [1]. The
use of performance indicators to describe the technical and
tactical aspects of game play form a crucial component of
this [4][7].
The technological innovations of recent years - in particular, advances in the field of position tracking - present new
challenges when it comes to analysing and interpreting this
data. While some of these standard indicators can be easily
generated from the raw data, their usefulness for performance analysis should be considered with a certain amount
of scepticism [8]. Today, sport science agrees that the key
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
2015 KDD Workshop on Large-Scale Sports Analytics , Sydney, Australia
Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00.
lies in using intelligent algorithms in order to derive complex performance indicators from the raw data that add real
value when it comes to game analysis [2, 3]. In this context,
the paper describes and evaluates an approach that enables
three types of individual ball possession to be detected using ball and player coordinate information. Based on this
we proposes a performance indicator called invasion index
for describing defensive and offensive success.
2.
BALL POSSESSION TYPES
From a sports science perspective, ball possession is the
most commonly investigated performance indicator [8]. Its
relevance is easy to understand, since having control of the
ball is a fundamental prerequisite for being able to invade the
opposing team’s third of the pitch and score goals. Existing
research based exclusively on ball possession at the team
level [5, 6, 9], because up until now, such data has been
collected by the competition information providers (CIP)
solely on the basis ball possession changes between teams.
The reason for this reduction in complexity is that ball possession data is collected by human data loggers concurrent
with the game, and it would be too expensive to manually
record data on an individual player basis.
When it comes to the sport scientific definition of individual possession, there is a certain amount of flexibility with
regard to how interruptions in play are handled, in deciding
when ball possession phases start and finish and how different levels of ball control are interpreted. In our approach,
only time intervals during which the ball is in play are considered when determining ball possession. When the ball is
in play, one of the two teams always has team ball possession
and one of the players always has individual ball possession.
When individual ball possession switches between two players, it is assumed that the pass reflects the tactical intent of
the first player. This time interval is thus classified as ball
possession for the first player. In the following we use these
definitions:
1. Individual ball possession (IBP) begins the moment
a player is able to perform an action with the ball
(following an IBP of another player or a game interruption). It ends the moment IBP for another player
begins.
2. Individual Ball Action (IBA) of a player begins the
moment this player is able to perform an action with
the ball and had no IBA prior to this. It ends the
moment the player is unable to perform any further
action with the ball. This interval no longer includes
the time from the ball being passed on until the start
of IBP of the next player.
3. Individual Ball Control (IBC) for a player begins when
IBA for this player begins and he is able to control the
ball during the IBA. What matters here is whether
a player has the ball sufficiently under control that
he can consciously chose between several play options.
It ends the moment this particular IBA of the player
ends.
3. DETECTION OF BALL POSSESSION
Automatic detection of ball possession involves a multistep process. The first step involves pre-processing the raw
data in order to reduce the number of unrealistic position
jumps and inaccuracies in the player and ball coordinates.
The heart of the procedure involves detecting the points
where the IBPs and IBAs start and the IBAs finishes. After
that, the level of ball control within the IBAs is estimated using a Bayesian network. The following subsections describe
details of the method.
3.1 IBP & IBA Starting Point
The IBP or IBA starting point is the moment when a
player starts to interact with the ball. This starts as soon
as the distance between the player and the ball falls below a threshold value and that player is nearest to the ball.
If the ball passes above a player, this could easily lead to
possession being incorrectly attributed, because of missing
z-coordinate. We obviate this problem by applying not only
a 2d separation threshold, but testing to check if a player
has interacted with the ball. We use the local maxima of
the amplitude of changes in ball velocity (ball accelerations)
together with narrow tolerance to filter out noise. We refer
to this as ball prediction.
3.2 IBA Endpoint
While the IBP endpoints are known once the IBA starting points have been identified, the IBA endpoints must be
specifically identified. There are three possible ways an IBA
for a player can end: 1) the game is interrupted, 2) another
player gains IBA or 3) the player is no longer able to interact with the ball. Whereas the first two cases are trivial
to detect, the last one requires special treatment. This involves checking whether a player will be still able to interact
with the ball within a certain time span. This is true if,
for example, a player kicks the ball a few metres ahead of
himself while dribbling, but not after making a pass or taking a shot at goal. Our approach make use of the current
positions and velocities of the ball to give an estimate of
their future locations. As long as it is possible for the player
currently in possession of the ball to control the ball within
one second, he will retain ball possession. We refer to this
as ball prediction.
3.3 Classification of IBC
Once the IBA starting and endpoints have been detected,
the IBA intervals can also be derived. The central question
is therefore which of the intervals represents a segment that
features ball control. This obviously depends upon a great
number of factors that may not be possible to ascertain from
positional information alone. For this reason, we decided to
use a Bayesian network to classify ball control based on a set
of variables like duration of ball action, average ball velocity
and acceleration, variance of ball velocity and acceleration,
average distance between the ball and the player in possession of the ball and number of opposing players within
certain distances.
4.
EVALUATION
A match between two teams in a top European league
served as the test sample in order to evaluate the quality
of IBP, IBA, IBC detection. Therefore an annotation with
reference data (ground truth), which was manually logged
by a trained, independent observer post game, formed the
basis for the evaluation. Table 1 shows the rates of IBP and
IBA detection for all frames of the game (complete data),
as well as within the frame sequences for which no tracking
errors were detected (flawless data). The parameters PRECISION and RECALL were used to assess the quality of the
recognition system, as it is used usual in machine learning.
RECALL is the relationship between the correctly identified changes and the total number of changes in the data
whereas PRECISION is the relationship between the correctly identified changes and the total number of changes in
the ground truth. Both parameters were calculated using a
0.6 s tolerance window.
Table 1: IBP and IBA detection results (in %)
Type
RECALL PRECISION
IBP
flawless data
80,1
86,1
complete data
78,0
76,9
IBA
flawless data
80,1
86,1
complete data
78,0
76,9
The recognition quality of IBCs for a given IBA interval was also determined by a comparison with the ground
truth. The degree of consistency according to Cohen is
κ=.39. With only 25 non-IBC intervals, the training set
for differentiating between IBA and IBC is very small, however. Moreover, the inter-rater reliability test between two
human observers on a subset of the intervals (n=98) did not
show complete consistency either (κ=.72). This indicates
that it may not be possible to fully objectify the ball control
construct.
5.
APPLICATION
Standard indicators are often not valid to performance [8]
and are unsuited to analyse e.g. the success of tactical patterns, substitutions or interventions in training, cause they
depend on chance and single case [7]. In order to describe
tactical subgoals rather than simple events, we proposes the
concept of invasions, which is based on individual ball control.
An invasion is an action where a player enters a defined
zone on the pitch with IBC. We divided the offensive half
into 5 zones as it is shown in fig. 1. Every zone gets a
weight which represents the increasing danger for scoring a
goal in this zone. The weight was empirically determined
by the scoring probability in this zone and transition probability between the zones in 42 matches of First German
Figure 2: Change of team performance in the course of a soccer game. Performance was calculated using
invasions to the offensive pitch half
Figure 1: Zones and weights used for calculating
team performances
or shots on goal, can be deduced directly from the individual ball possession data. These events are logged by CIPs
as default, however, due to the manual data collection process the event lists are not always complete and the events’
time stamps are sometimes imprecise. Automatic detection
methods can thus help to ensure the quality of match data
and potentially reduce the loggers’ workload.
In addition, being able to detect ball possession is a fundamental prerequisite for discerning higher value tactical structures like availability, pressing strategies or attacking/defending
performance. The capability to recognise ball possession
types holds considerable potential for improving the quality of match analysis in professional soccer. The concept of
invasions provides a first example for this.
7.
Professional Soccer League (Bundesliga). In order to quantify attacking and defending performance we calculate an
invasion index, by summing up the weight of the highest
zone reach during every team ball possession phase.
As an example, fig. 2 shows the invasion index over the
course of one match of Bundesliga (mean value of 5 min intervals). After the red card in minute 35, there was a breakdown in the offensive performance of the home team (red
line). Nevertheless this team scored two gaols afterwards.
From this point in time the away team (blue line) was able
to realize much more invasions compared to the home team,
but was not able to score a goal. In the last 5 minutes the
home team forgoes on any offensive activities for saving the
result. However, indicators like ball possession, passing accuracy or duels won do not show the dominance of the away
team. This example shows how the result and also standard performance indicators can hide the real performance
of teams.
6. CONCLUSION
The quality of tracking is at a high enough level today
to allow for individual possession to be reliably detected.
Using the method on uncorrected data results in flaws at
the raw data level and detection errors accumulating that
leads in turn to a cumulative detection accuracy of around
80%. However, tracking quality will surely increase in the
next years due to technological progress.
This positive outcome suggests that the method can be
used for a wide variety of potential applications. For example information about basic events, such as passes, tackles,
REFERENCES
[1] C. Carling, T. Reilly, and A. M. Williams. Performance
assessment for field sports: physiological, and match
notational assessment in practice. Routledge, 2009.
[2] A. Grunz, D. Memmert, and J. Perl. Tactical pattern
recognition in soccer games by means of special
self-organizing maps. Human movement science,
31(2):334–343, 2012.
[3] J. Gudmundsson and T. Wolle. Football analysis using
spatio-temporal tools. Computers, Environment and
Urban Systems, 47:16–27, 2014.
[4] M. Hughes and R. Bartlett. Possession as a
performance indicator in soccer. Journal of Sport
Science, 20(10):739–754, 2001.
[5] M. Hughes and I. Franks. Analysis of passing
sequences, shots and goals in soccer. Journal of Sports
Sciences, 23(5):509–514, 2005.
[6] P. Jones, N. James, and S. D. Mellalieu. Possession as a
performance indicator in soccer. International Journal
of Performance Analysis in Sport, 4(1):98–102, 2004.
[7] M. Lames and T. McGarry. Performance analysis
research: Meeting the challenge. International Journal
of Performance Analysis in Sport, 7(1):62–79, 2007.
[8] R. Mackenzie and C. Cushion. Performance analysis in
football: A critical review and implications for future
research. Journal of sports sciences, 31(6):639–676,
2013.
[9] J. Pratas, A. Volossovitch, and A. Ferreira. The effect
of situational variables on teams performance in
offensive sequences ending in a shot on goal. a case
study. Open Sports Sciences Journal, 5:193–199, 2012.