The performance analysis of a soccer team has become an important topic for soccer coaches. Parameters like the number of shots, passes or sprints during a match provides information about the game quality. However, currently available systems are based on cost-expensive video analysis, which requires a pre-installed infrastructure. Consequently, such systems are only open to professional teams. The use of low-cost wearables represents an alternative to make such performance analysis accessible to hobby teams. This paper focuses on the evaluation of diferent Machine Learning (ML) approaches for the classification of simple, soccer-specific activities (such as standing, walking, running, passing and shooting) based on Micro-Electro-Mechanical Systems (MEMS) accelerometers. To do so, the sensors are mounted on the soccer player's shin guards. Diverse ML algorithms as well as diferent dimensionality reduction algorithms are investigated. The best approach shows a macro-precision score of 97% and a macro-recall score of 96%. The final goal is to develop smart shin guards, which can georeference soccer-specific activities in conjunction with sports statistics. The positioning system is based on Global Navigation Satellite Systems (GNSS). This scientific study is part of the Austrian Space Applications Programme (ASAP) 15 and funded by the Federal Ministry of Transport, Innovation and Technology via the Austrian Research Promotion Agency (FFG).
Observing the activities of each soccer player in a match is a big challenge for coaches.
For instance, the number of passes, shots, sprints done by the players as well as their
positions on the field represent a valued information for the team. The evaluation of
those parameters can be used to improve existing strategies or develop new ones that
may derive a benefit concerning the team. However, currently available systems are
based on video analysis. skills.lab [
Being in an upright position with both feet on the ground.
Moving on foot at moderate speed (≈ 5.5 km/h).
Moving on foot at advanced speed (> 8 km/h).
Player tries to kick the ball to another teammate.
A more intense kick.
for soccer players. Specific game situations, utilizing state-of-the-art measurement
technologies, can be reproduced to improve the skills of the soccer players [
Commercial wearables are cost-efective to observe the activity and fitness of persons.
Smart gadgets, such as smartwatches, are gaining popularity; the number of wearables
sold is expected to be around 630 million worldwide until 2024 [
This paper deals with the first development steps and examinations of smart shin guards for soccer players. GNSS-sensors are utilized to get the player’s position on the ifeld. Micro-Electro-Mechanical Systems (MEMS) accelerometers with Machine Learning (ML) methodologies are used for activity recognition. The result is a georeferenced activity that enables performance analysis. The focus will be on evaluating diferent ML approaches for the classification of simple, soccer-specified activities. The activities and their definition are listed in Table 1. Among other things, the concept concerning the GNSS-based positioning system will be shown.
In total, four diferent ML algorithms are investigated: Logistic Regression, Support Vector Machine (SVM), Random Forest Classifier and an Articfiial Neural Network (ANN) in form of a Multilayer Perceptron (MLP). Features are chosen in the timeand wavelet-domain. Furthermore, it has been examined, whether the Principal Component Analysis (PCA) or the Linear Discriminant Analysis (LDA) in the meaning of dimensionality reduction algorithms can improve the classification process or not.
Triaxial accelerometers are popular sensors for activity recognition (AR). The authors
of [
The MPU9250 MotionTracking device [12], produced by InvenSense Inc., is chosen as the embedded Inertial Measurement Unit (IMU). The MPU-9250 consists of triaxial acceleromters, a triaxial gyroscopes and a triaxial magnetometer. This sensor is based on MEMS technology. However, in this study only the accelerometer data is used for AR. The range of the accelerometer is set to its maximum, namely ± 16 g (1 g ≈ 9.807 m/s2 at latitude of Graz). The output rate was set to 100 Hz.
The generation of the GNSS position of the soccer player is done via the Neo-M8T chip from u-blox. This chip supports the raw data output so that a position, using an own software, can be calculated. For the first investigations, two u-blox evaluation kits (left and right shin guard) are used. The appropriate antennas are still under investigations for this use case. Therefore, the antennas from u-blox are utilized, which are included in the evaluation kits. The GNSS data recording was done with a notebook in the backpack of the player. With these two raw GNSS data sets, a combined Multi-GNSS position with an update rate of 1 Hz can be calculated.
The provisional test setup, including the sensor orientation of the IMU, is shown in Figure 1. Each shin guard will be finally equipped with one IMU, one GNSS chip, one GNSS antenna and the necessary processor units. The IMU is mounted at the lower end of the shin guard to sense the motion performed during a kick as well as possible. The two IMUs are synchronized based on the Precision Time Protocol (PTP). The communication between the two IMUs takes place via Wi-Fi. In this rfist approach, the recorded accelerometer data is stored on an SD card.
Before the data collection started, the MEMS accelerometer was calibrated. In the calibration process the bias and scale factor of each axis are determined. The accelerometer M 92 PU 50 z y y z x (a) Hardware Components. Photograph by
Stefan Laller. (b) Sensor Orientation (IMU). Sketch by
Karin Mascher [
For segmentation, a window size W of 256 samples (2.56 s) was chosen. Standing,
walking and running were segmented with an overlapping window with the size W and
an overlap of 50 %. Kicks were cut out so that the acceleration peak is in the center
of the window W . Labels were manually assigned [
This section deals with concept of the AR strategy. Based on the labeled data, different ML approaches have been investigated, which will be explained in more detail below. Computations were done in Python 3.7 using the ML modules scikit-learn and TensorFlow2 as well as the wavelet transform software PyWavelets 1.1.1.
Features are also chosen in the wavelet-domain. Therefore, the Discrete Wavelet
Transform (DWT) is utilized as a filter bank. The implementation of the DWT as a filter bank
can be seen as a cascade of high-pass (signal details) and low-pass (smoothing efect)
iflters [14]. The signal is decomposed into so-called approximation and detail coeficients,
which corresponds to diferent sub-frequency bands. The result is a list of coeficients
arrays [15]: approximation coeficients array (cA M ) and the detail coeficients arrays
(cDM , . . . , cD1) (M ≥ 0 refers to the maximum level of decomposition). Depending
on the chosen wavelet, the maximum level of decomposition varies. However, the
frequency ranges corresponding to diferent decomposition levels for this special case are
listed in Table 2 [
The usage of the DWT gives a compact representation of the energy distribution in frequency- and time-domain [14]. As an example, Figure 2 shows the wavelet analysis of two shot records. The signal is expressed in terms of the total acceleration of the shooting leg. The chosen wavelet is the reverse biorthogonal 3.1 (rbio3.1) wavelet. Similar analysis were performed for the other activities.
1. Removing features with low variance 2. Removing features with low variance followed by Principal Component Analysis (PCA): variance explained ≥ 97.5% (keep 97.5% of the “variability” of the 50 60 70 80
original data set) (cf. Section 5.5) 3. Removing features with low variance followed by Linear Discriminant Analysis (LDA) (cf. Section 5.5) Those transformation values are applied to the test set.
The box “Machine Learning” pictures the training of the diferent ML algorithms based on the diferent feature subsets. In total, four diferent ML algorithms are investigated: 1. Logistic Regression 2. Support Vector Machine (SVM) 3. Random Forest Classifier 4. Artificial Neural Network (ANN): Multilayer Perceptron (MLP) The training comprised the selection of the model parameters as well as the hyperparameter tuning. Since the input data is imbalanced, small classes receive respectively stronger weights. 5-fold cross-validation was used for model evaluation (hyper-parameter tuning, feature selection, ...). In a final step, the test set is fed into the trained model. The output are predicted labels that are compared with the actual labels. Hence, an unbiased quality of the ML model can be assessed.
To sum up: the performances of four diferent ML algorithms that are based on three diferent feature selection methods are analyzed (12 combinations in total).
The AR strategy is split into two parts [
Features are chosen in the time- and wavelet-domain and for each classifier separately.
Wavelet-based features are expressed in the terms that are used in Table 2. The
appropriate mother wavelet is visually selected based on the resemblance between the signal
of interest and the diferent wavelet types [
The total acceleration atotal is calculated as follows
atotal,i = √︂xi2 + yi2 + zi2,
(1)
where i is the ith element of the acceleration vectors x, y and z of dimension W . W refers
to the window size. The largest and second largest value from atotal are selected as
features. These features imply the intensity of a kick as well as the corresponding
preand post-impacts [
The Signal Magnitude Area (SMA) (introduced by [
SMA = where cDj,i (j ∈ {2, . . . , M }) is associated with the ith component of the detail coeficients array. Nj is the dimension of the coeficients array at level j and M refers to the maximum level of decomposition. The analysis is based on the rbio3.1 wavelet. cD1 is not considered, since frequencies between 25 Hz to 50 Hz are of no interest to the preclassification process. The mean and maximum of absolute approximation coeficients array cAM as well as the RMS provide information about the low frequency components. The mean is obtained using the following formula:
MEAN(cAM ) =
N 1 ∑︂ |cAM,i|, N i=1 where N is the dimension of the approximation coeficients array. The formula for the RMS and maximum are analog to those of the detail coeficients. Here also the rbio3.1 wavelet is used.
The maximum, the mean (Formula 5) and the Interquartile Range (IQR) (Formula 6) are computed for each sensor axis and chosen as features in the time-domain. The following formulas are illustrated for an arbitrary vector a of dimension W . where ai is the ith element of the vector a.
MEAN(a) =
W 1 ∑︂ ai, W
i=1
IQR(a) = Q3(a) − Q1(a), where Q1(a) is the first quantile and
Q3(a) the third quantile of the vector a.
(2)
(3)
(4)
(5)
(6)
The total SMA [
Studies [
The two largest values of the total acceleration of the xz-plane are used as a feature.The Pearson correlation coeficient between the total acceleration (Equation 1) of the event and supporting leg is computed and shown in the following formula: r = √︃ ∑︁iW=1 (︂ a(tsoutaplp,iort) − a¯(tsoutaplport))︂ (︂
(event) (event))︂
atotal,i − a¯total
∑︁iW=1 (︂ at(soutaplp,iort) − a¯(tsoutaplport))︂ 2√︃∑︁iW=1 (︂ at(eovtaeln,ti) − a¯t(eovtaelnt))︂ 2
(8)
(event) and a¯t(soutaplport) represent the mean values of the event leg and the supporting
where a¯total
leg, respectively. This feature gives information about the activity of the supporting leg.
Since during the event of a shot, the supporting leg is more active than during a pass,
a correlation between the activity of the right and left leg is given [
Due to the curse of dimensionality that says “as the number of variables under
consideration increases, the number of possible solutions also increases, but exponentially”
(William S. Noble [17, p.1567]), it is challenging for ML algorithms to find an accurate
solution. Hence, one solution is to reduce the number of features by projecting the
original feature space onto a lower-dimensional one and that with a minimum of information
loss [
Two dimensionality reduction algorithms are investigated in the course of this study: 1. Principal Component Analysis (PCA)
PCA is an unsupervised dimensionality reduction algorithm. The algorithm tries to
ifnd a hyperplane that preserves the maximum variance of the original. Correlated
features are transformed in linearly uncorrelated ones. The new features are named
principal components (PCs) [
LDA is a supervised dimensionality reduction algorithm, which aims to find a
hyperplane that keeps the classes as far away as possible. The new features are
known as linear discriminants (LDs) [
More information about those dimensionality reduction algorithms can be found in [19, 20, 21, 22]. Table 3 now contains the total number of features based on the diferent feature selection methods.
The chosen performance metrics are based on the macro-averages of the precision and recall scores. The macro-average is chosen since this metric is preferred when dealing with imbalanced data sets.
The training and test scores for all models were above 96.9% (precision) and 97.5% (recall). All approaches show potential to correctly classify standing, walking, running, passes and shots. The nfial models are trained on the whole data set and applied to real, but simple data to find out which model is best able to detect the activities. That data set is composed of five sequences that include the soccer-specific gestures. Each sequence consists of: standing (≈ 1 min), followed by running with increasing speed ( 8 km/h up to 12 km/h), a short pause and a walking phase (≈ 4 km/h). During the running phase, two passes and one shot is performed. During the walking phase, one pass is done. Figure 5 illustrates the validation set expressed in form of the total acceleration. Kicks are performed with the left and right leg as well as with diferent techniques (insight foot, outside foot, full instep).
The used window size is 256 samples. The overlap is around 60%. Tables 4 to 6 show the precision and recall scores of the diferent ML models (Section 5.2).
The best results show Logistic Regression and the Random Forest Classifier used in combination with the Linear Discriminant Analysis (LDA). They are the only models that received perfect precision and recall scores for the classes pass and shot. Misclassiifcations only took place in the transition areas between standing, walking and running.
left leg pe ac Hence, those errors are not considered as severe ones. The worst performance shows the Random Forest Classifier in combination with the Principal Component Analysis (PCA). The recall score for walking, for example, was only around 50%. Since Decision Tree classifiers use linear and orthogonal decision boundaries, they tend to overfit and do not work well an new instances that are slightly diferent from the training data. In this special case, the walking phase in the real data set is performed at an average speed of 4 km/h, but the walking instances used for training are done at an average speed of 5.5 km/h.
This study compared diferent ML models with each other based on a modular classiifcation scheme. The performance was evaluated on real, but simple data to figure out which ML model performs the best. First, the use of two IMUs is a powerful tool to distinguish passes from shots. On the one hand, the intensity of the kick is sensed by the event leg. On the other hand, the activity of supporting leg is higher during a shot than during a pass. Secondly, it has been shown that the Linear Discriminant Analysis (LDA) is well suited for the selection of an appropriate feature subset. LDA significantly outperforms the two other feature selection methods and worked pretty well for all ML algorithms. However, applying the Principal Component Analysis (PCA) to the feature subset does not improve the classification process. Third, the ANN only showed a mediocre performance. One explanation could be that Neural Networks are designed to deal with large data sets. The existing data set is relatively small. Fourth, the sensor exceeded its maximal range of ± 16 g when performing a shot. But it was suficient to separate passes from shots.
The used data is only based on one test person. Thus, future studies will aim to collect data from a suficient number of test persons to create a model that can be used for a broader audience. Another goal is to implement more activities in the classification process, such as tackling, dribbling and other more complex gestures. The activities still need to be georeferenced to get a Location-based Service (LBS) that enables an easier performance analysis of soccer players. The concept is shown in the last section of this paper (Section 8).
To sum it up, under laboratory conditions, it is possible to detect simple, soccer specified activities using a MEMS accelerometer. The best performance has been achieved with Logistic Regression and the Random Forest Classifier combined with the Linear Discriminant Analysis (LDA). These models had a macro-precision score of 97.4% and a macro-recall score of 96.1%. They were also able to detect all passes and shots.
For the processing of the GNSS raw data the software “eRTK” from the Institute of Geodesy (Working Group Navigation) is used. This software uses GPS and GALILEO observations for the determination of a Multi-GNSS Single Point Positioning (SPP). The positions from both shin guards are combined with an algorithm, which weights the positions with the covariance matrix from the least-squares adjustment [23]. The idea for using Multi-Receiver and Multi-GNSS is that a more robust position can be calculated. The verification of the position solution is done by high-priced GNSS equipment. In a further step, the time synchronization of the position and the detected activity is done. Due to the GPS timestamp of the player’s position and a Coordinated Universal Time (UTC) timestamp of the activity, this task can be done easily. The result is a georeferenced activity for performance analysis of soccer player.
Further investigations will show which accuracy and availability can be achieved with this setup. Furthermore, it has to be investigated if the whole concept can be used in a real soccer game and can be used for performance analysis of soccer players.
This study is funded by the Austrian Federal Ministry of Transport, Innovation and Technology via the Austrian Research Promotion Agency (FFG) as part of the programme ASAP 15. The whole consortium consists of c.c.com Moser GmbH (Grambach, AUSTRIA), the Institute of Geodesy at Graz University of Technology (Graz, AUSTRIA) and the Institute of Human Movement Science, Sport and Health at University of Graz (Graz, AUSTRIA). Thanks to all for their support and cooperation! Big thanks to Christoph Schmied for the technical support. Thanks to Bernd Mo¨lg, who built the custom-made shin guards. Both Christoph and Bernd are members of the Institute of Geodesy.