Towards an automatic detection system of sports talents: an approach to Tae Kwon Do

Hacia un sistema de detección automática de talento deportivo: una aplicación al Tae Kwon Do

Em direção a um sistema de detecção automática de talento esportivo: uma aplicação para o Taekwondo

This work is licensed under Creative Commons Attribution 4.0 International.

The use of Machine Learning [ML] theory has been extended to several emerging areas of study, such as data security and commerce, among others (Trejo & Miramá, 2018; Urcuqui & Navarro, 2016; Vergara, Martínez & Caicedo, 2017). In such a way, sport is effectively combined with ML Theory due to the large amount of data that can be extracted from a singular athlete or team. In this sense, ML is one of the most used theories for analysis in sports, which has focused on sport performance (Alderson, 2015), diagnosis of sport injuries (Zelic, Kononenko, Lavrac, & Vuga, 1997), and forecasting sport results (Valero, 2017).

Tae Kwon Do is a well known Korean martial art and Olympic combat sport which stands out for the variety and the impressive of its kicking techniques. In such a sense, several proposes have been developed to improve the competitors training by using a motion system with body and visual sensors and ML for analysis (Kwon & Gross, 2005). A hybrid approach sensing technique in conjunction with Hidden Markov Model [HMM] (Kwon, 2013) and a humanoid robot able to interact with athletes, in order to give instructions and improve training (Muscolo & Recchiuto, 2016). In the mean time, there exists works focused on the athlete to analyze the complex techniques in contact sports by using video frames and Deep Learning [DL] for predicting the action to be executed (Kong, Wei, & Huang, 2018). Furthermore, an approach to develop a dynamic evaluation of Tae Kwon Do by using the classification method of Genetic Algorithms with Support Vector Machine [GA-SVM] was proposed by Zhong, Hung, Yang, and Huang (2016) .

However, experts consider there is a gap in the process of athlete selection according to their expectations and reality, and the best of our knowledge, no principled studies have been conducted to recognize athletes in Tae Kwon Do and identify the main features of athletes with high competitive performance. For this reason, the aim of this paper is to develop a classification based system for determining the key features for identifying athletes towards a high performance in Tae Kwon Do; in order to fill out this paper, we apply feature selection and classification stages to data provided by the Federación Ecuatoriana de Tae Kwon Do [FETKD]. For the former stage, we propose to use wrapper and embedded methods, as long as for the next stage, supervised classification was considered, by using two well-known algorithms such as Decision Tree [DT] and Support Vector Machine [SVM]. This approach could allow us to make decisions –with reliable results as possible– about athletes suitability with major expectation of high performance.

The main contribution of this work is to provide a support system for the athlete selection based on exports opinion whereby we can identify the best candidates. As well as the extraction of the key features, which can use for specific training oriented to improve the weak features provided by the support system. This work is according to lead an high performance athlete –firstly towards Olympic games and several events where national team could participate, locally or internationally–, from an early age.

The rest of this paper is organized as follows: in section II we define materials and methods for feature set and pre-processing, feature selection and classifiers, which are evaluated by their performance; in section III we show the experimental result and in section IV the conclusions and discussion obtained from this research.

In this section we detail feature set and pre-processing, feature selection and classifiers, and metrics to evaluate its performance, as depicted in Figure 1. About feature set design, this has been proposed and extracted by experts based on their experiences from athletes selection area at FETKD. Furthermore, in the feature selection stage, we have introduced the use of wrapper–embedded methods in conjunction with SVM and DT, respectively, as classification algorithms. Filter, embedded and wrapper methods are currently used to select the best set of features. For this reason, we have chosen a wrapper method based on Recursive Feature Elimination [RFE], which could achieve high classification Accuracy (A). Meanwhile, an embedded method uses feature selection and classifiers in conjunction for learning key features, which contribute to improve A and avoid over-fitting (Liu, Wang, Zhao, Shen, & Konan, 2017; Blum & Langley, 1997; Langley, 1994). We have dismissed the use of filters due to proposed wrapper and embedded methods outperforms filter algorithms (Suto, Oniga, & Sitar, 2016). Whereas the classification, we have used a well-known algorithm –DT–, because it closely resembles human reasoning and presents a simple hierarchical structure for the user understanding and decision making (Kotsiantis, 2013; Badr, Abdelkarim, Hanane, & Mohammed, 2014). Besides, we have used SVM due to it provides a higher A, and it is, for this reason, one of the most powerful techniques of ML which has been proven to be a robust algorithm that generalizes well into real life engineering applications for forecasting (Parikh & Shah, 2016; Shi, Duan, Ma, & Weng, 2012; Zhang, 2012). The ML algorithms must be assessed its performance, consequently we have chosen metrics associated to those algorithms (Lara, 2015). We have used Matlab®R2016a, a PC re(TM) i7-5500U with 2.4–2.39 GHz and 8GB of RAM for development of the experiment.

Figure 1
Block diagram for the proposed system

A. Feature Set and Pre-Processing Stage

An sporting talent is an athlete who possess main features required to get a higher probability of consolidation in a sport. By this way, the traditional models of athletes selection are based on the ascription to a certain activity. It can be described two models to consider, which are the empirical or scientific model and the formative or development model (Brotons, 2005). The process to search and identify athletes potentially successful related to Tae Kwon Do is based on a mix of selection models previously mentioned. Experts have developed a recognition process based on well-defined features such as gender, category, weight and overweight, which are related to his/her physic somatotype. In addition, several tests have been developed to obtain some features like: physical abilities and techniques-tactics abilities, which are related to sports adaptation. For our understanding, gender feature determines if the candidate is male or female, while category places an athlete on their respective weight, age and gender established by the World Tae Kwon Do as described in Table 1.

Table 1
Relation between weight, age and gender feature to category

The relationship between these features is understandable in a way that an athlete belongs to a category, which is limited by the higher and lower weight. The athlete could be located in three possible cases; under, into, and over the limit; the overweight has a positive and negative relation concerning the facts mentioned previously; physical and technical-tactics abilities has a subdivision related to training stage or orientation that is necessary to work into the sporting process. Physical abilities are associated to the athlete capacities as strength, speed, endurance, flexibility and coordination, while, technical-tactics abilities allow us to adopt the train stage and evaluate as necessary for coaches, conditioned by specific sport needs. We developed pre-processing in a general context to enhance the discrimination in all features on our data set, by eliminating the lineal trend and label all the features to be used. Over our case we removed lineal trend by using zero mean and variance equal to one (µ=0, ν=1), which allows improving visualization of our feature set in the same range; on the other hand, feature set is labeled such as gender (X1), category (X2 ), weight (X3 ), overweight (X4 ), physical abilities (X5 ) and technical-tactical abilities (X6 ).

B. Feature Selection Stage and Classifiers

Feature selection stage was developed to identify the principal sets of relevant features from athletes, which will enable us to determine the main features to work toward high competitive athlete performance. It is performed a benchmark study of two most used feature selection methods which are named “embedded” and “wrapper”. The goal of this methods it to obtain matrices which provide most of the discriminative information to classify the athlete, while avoiding overfitting. By using the embedded method, in this work is necessary to select an algorithm, which as the main criterion uses Mutual Information [MI] between feature x and the output y, as follows in (1).

(1)

where, the marginal entropy is defined as H(y), while conditional entropy is associated with H(y|x) between output y and feature set x, through generating an iteratively builds by dividing the data taking advantage according to its importance for the classification task. The algorithm used is DT, which is considered a non-parametric supervised learning algorithm and is principally used for both, classification and regression problems. The goal of this algorithm is oriented towards a model which can predict the value of a variable by learning decision rules inferred from the data features. The free parameter of DT algorithm is the depth or leafiness and it has to be adjusted in order to maximize the classification performance, avoiding overfitting to the training set. The tree is shaped by a root node, internal nodes and terminal nodes; moreover, in each node a rule is established, which is the entrusted to produce the binary selection extend to the final node which represents a class. All the possible branches are dependents to each node values can take. In this way, the algorithm generates sequential decisions to predict values, as of representative features of the data. Introduces an approach based in information theory, where the choice of a feature it is directly related with entropy, which is described as a measure in a system uncertainty that allows us to know the necessary average amount of bits can be adapted to the output of the algorithm. This parameter is represented by Equation 2

(2)

where, C describes a set of the class which may belong to such an example and pi is the likelihood that given example belong to i-th class. For a wrapper method, we used RFE, which has a base on a backward elimination method. Their operation is based on iteratively removing features from data, seeking to choose the features which lead to the largest margin of class separation by using SVM as a classifier. In our case, the selected was ν-SVM, enabling the variation necessary of a free parameter ν which control the number of support vectors. The ν-SVM algorithm is defined in summary as follows (see Schölkopf & Smola (2002) for details. By using a labeled training data set (3):

(3)

donde (4)

(4)

and (5)

(5)

and given a nonlinear mapping the ν-SVM methods solves (6).

(6)

subject to (7) and (8)

(7)

(8)

where, w and b define a linear classifier in the feature space and the positive slack variables enabling to deal with errors, it is associated with ξi . It should be taken that the appropriate choice of nonlinear mapping θ allows us guarantees that the transformed samples hold a major probability for being linearly separable in the feature space. In this context, we can identify that the variable is controlled through coefficient, which provides a new degree of freedom to the margin. Furthermore, the size of the margin increasing linearly with the variation of the parameter ρ. Therefore, adjusting ν in the range [0;1] in the ν-SVM algorithm allows performing the trade-off between the training error and the generalization error, which is defined as an upper bound on the fraction of margin errors and is also a lower bound on the fraction of support vectors. The optimal solution of the primal problem (6) could be obtained by using its dual problem counterpart, introducing (9)

(9)

while decision function for any text vector x* is finally outlined by (10).

(10)

It is possible to describe constraints in (6) as Lagrange multipliers defined by αi , being the Support Vectors [SV] those training samples xi with non-zero Lagrange multipliers αi ≠ 0; and the bias term b calculated by using the unbounded Lagrange multipliers as (11).

(11)

where k is the number of unbounded Lagrange multipliers (0 < αi < C). SVM present a particularity around the decision function f(x), defined as a function of a small subset of the training examples described as the support vectors. Those are examples closes to the decision boundary and lie on the margin as well as those wrong-class examples. The existence of such support vectors is at the origin of the computational properties SVM and their competitive classification performance (see Guyon, Weston, Barnhill, and Vapnik, 2002) for more details about the SVM algorithm related to linear and non-linear.

C. Performance

This stage was developed to evaluate the classification performance. By performing the determination of the athletes labeled with the value “1”, equal to an suitable candidate, and “-1” for a not suitable candidate; an information collation is carried out by generated and real labels. The established measures for performance of classifiers –Accuracy (A), Precision (P), Sensitivity (R), Specificity (S) and Balanced Error Rate (BER)– are described with the equations 12, 13, 14, 15 and 16, respectively.

(12)

(13)

(14)

(15)

(16)

where: NC belongs to the number of patterns correctly classified; NT make reference to the number of the used patterns in the classification; NTP is the number of true positives; NFP is the number of false positives; NTN express the number of true negatives; and NFN the number of false negative.

We calculated these performance measures for each validation used in all the cases proposed.

The results obtained throughout this research allows us to perform an analysis and approach toward a high-performance athlete, following the identification of main features and the athlete classification. The data to be analyzed corresponds to a total of 76 athletes, which was divided into two groups. The first group be owned by the training set of the algorithms, with a total of 54 athletes. While the test set has 22 athletes, these last are the most recent obtained in 2018.

In other words, our training set is equivalent to 71.052%, while the test set is 28.948%. This will allow us to carry out in a feasible way the feature selection and athletes detection. Enabling reliable results from the supervised algorithms and avoiding over-fitting.

By making use of a three-dimensional plane, in Figure 2a we can see the feature set surface provided, while Figure 2b presents to us the feature set surface after pre-processing. This stage works in the sense of removing lineal trend and place all the feature set on the same range, by using of µ=0 and ν=1.

Figure 2a
Surface representation of original feature set (a)

Figure 2b
Surface representation of pre-processing feature set

The original output of athletes classification is depicted in Figure 3, which allows to compare with the output delivered by DT and SVM algorithms, described below, where, as mentioned, a value “1” is assigned to a suitable candidate, while “-1” corresponds to a non suitable candidate.

Figure 3
Original output of athletes classification

A. Results Using DT

DT algorithm obtained a model for the input matrix by using the features established previously and the output matrix corresponding to athletes classification. It made possible to induce a tree, as depicted in Figure 4, which chosen three key features. Beginning from the top node with the rule X5 ≥ 0.827944, followed by X4 ≥ 0.366235 and finally X6 ≥ −0.0302377, which made possible to classify into any 1 of the 4 possible leafs.

Figure 4
Tree representation considering main features, by using DT algorithm

This representation establishes different thresholds depending amplitude that would make it possible to identify the eligibility of an athlete. Where was determined a root node X5, in this way if threshold value it is exceeded the athlete is considered suitable, in the case of a candidate not succeeding this value proceed to take of a new decision. The next node for decision making is X4, will no have to go over the threshold value of 0.36, so that the athlete is not discarded, allowing a concatenation with the last feature X6, which could be greater than or equal to -0.03, so that the candidate be suitable. By using all the feature set for a supervised classification provides the following performance measures: A%=86.3636, P%=90, R%=81.8182, S%=90.9091 and BER=0.1364. The output of athletes classification delivered by the algorithm is depicted in Figure 5. Using main features proposed by DT algorithm, we determine that performance parameters and classification, which are the same in the case of use all features.

Figure 5
DT output of athletes classification

B. Results Using SVM

RFE algorithm obtained a model for the input matrix by using the features established previously and the output matrix for feature selection. In this sense, the feature selection framework was made based on SVM-RFE algorithms, which identify the key features from the main one toward less important feature, based on the weights of each one. For our case, through the use of this method, we determine three key features (described in Table 2), which are translated in percentage according to X4=70.820%, X6=22.289% and X2=6.891%.

Table 2
Main features delivered by using SVM-RFE algorithm

We decided to use in classification algorithms three sets of features, which are the key sets provided by DT, RFE and all features. Furthermore, for classification we use two different kernel in ν-SVM algorithm, which are Lineal and Radial Basis Function [RBF] Kernel. The adjustment of parameter ν is carried out with a constant variation of 0.01 in the established range for the algorithm, those experimental results are described below. We detail performance measures obtained by using lineal kernel (see Table 3). Figure 6 depicts the output of athletes classification delivered by each algorithm proposed. Figure 6a represents the corresponding athletes classification provided by the ν-SVM lineal kernel algorithm, with DT feature set, where v=0.29; Figure 6b shows the output of the algorithm with ν=0.18 by using RFE feature set; and Figure 6c define an amount of ν=0.24 for all features.

Table 3
Performance measures by using ν-SVM lineal kernel and different features sets

Figure 6a
The output of athletes classification by using ν-SVM lineal kernel algorithm with DT feature set

Figure 6b
The output of athletes classification by using ν-SVM lineal kernel algorithm with RFE feature set

Figure 6c
The output of athletes classification by using ν-SVM lineal kernel algorithm with all features

For the RBF kernel case, Table 4 depicts the measures of performance delivered by the algorithm and Figure 7 shows the athlete classification. The features employed in these algorithms are the same used in lineal kernel algorithms. Figure 7a depicts the corresponding athletes classification provided by the ν-SVM RBF kernel algorithm by using DT feature set, where 0.96 is the better value of ν. Figure 7b shows the output of the algorithm with ν=0.11 by using RFE feature set and Figure 6c define an amount of ν=0.23 for all features.

Table 4

Performance measures by using ν-SVM RBF kernel and different features sets

Figure 7a
The output of athletes classification by using ν-SVM RBF kernel algorithm with DT feature set

Figure 7b
The output of athletes classification by using ν-SVM RBF kernel algorithm with RFE feature set

Figure 7c
The output of athletes classification by using ν-SVM RBF kernel algorithm with all features

It can also seen in Figures 6a, 6b, 7a and 7b, by using the same sets previously mentioned, that there is a noticeable similarity around athletes classification, providing X4 and X6 as common features. Nevertheless, athletes classification can be represented graphically with the use of a three-dimensional plane, enabling identify the separability delivered by the algorithms according to the used kernel, where the axes are assigned as (X4, X5, X6) to (X, Y, Z) and (X2, X4, X6) to (X, Y, Z) respectively. For our better knowledge, Figure 8 depicts the output classification by using lineal and RBF kernel, with DT and RFE sets, where (+) corresponds to a suitable athlete and (°) is a non suitable athlete. This representation makes hard to identify which are the main features and a great difference from one algorithm to another can’t be observed. Figure 8a and 8b depict the corresponding athletes classification provided by the ν-SVM lineal kernel algorithm by using DT and RFE features sets, respectively. Figure 8c and 8d show the output of the algorithm in a three dimensional plane by using DT and RFE feature set.

Figure 8a
The output of athletes classification in a three-dimensional plane by using ν-SVM corresponds to DT feature set with Lineal kernel

Figure 8b
The output of athletes classification in a three-dimensional plane by using ν-SVM corresponds to RFE feature set with Lineal kernel

Figure 8c
The output of athletes classification in a three-dimensional plane by using ν-SVM corresponds to DT feature set with RBF kernel

Figure 8d
The output of athletes classification in a three-dimensional plane by using ν-SVM corresponds to RFE feature set with RBF kernel

In this paper we have proposed a detection of sports talents by using machine learning theory oriented toward Tae Kwon Do. First, a novel yet straightforward method has offered for feature selection and classifiers for an objective and impartial selection of athletes, which has been solved by using embedded and wrapper methods associated to DT and SVM algorithms. The analyzed scenario corresponds to the last years data corresponding to athletes from the Ecuadorian Tae Kwon Do National Team. Second, we supplied the managers with additional information about the most relevant features to be taken into account, for this purpose, the features of athletes were measured up, giving a clearer view of which is the most critical of them on a new systematic and easy-to-handle representation with the coaches. Features analysis has been allowed to detect the best candidates and identify features will be the essential item to work on it. The use of supervised algorithms makes this support system more than an athlete classification tool and is solidly based on the analysis of well-defined features. The application of these two cases of study (different theory of ML) highlights the practical convenience and usability of this approach. In our results, the analysis of feature selection showing a reduction to the half of these in both cases. It is possible identify two common features delivered by the algorithms, which are X6 and X4. For the DT algorithm, the performance measures are: A%=86.3636, P%=90, R%=81.8182, S%=90.9091 and BER=0.1364. Through the use of ν-SVM algorithm with lineal kernel, we identify the best case, which uses ν=0.29 and key feature set delivered by DT, providing the same performance measures of DT algorithm, however, it outputs a different classification. Furthermore, for the best case of ν-SVM algorithm by using RBF kernel, we obtain a value of ν=0.23 and using all features set; it provides the next performance measures: A%=90.9091, P%=100, R%=81.8182, S%=100 and BER=0.0909. Thus, the best algorithm for athletes classification in our case is associated to ν-SVM with RBF kernel, which outputs a high-performance measures. We are able to concluded that the proposed novel support system can be useful to determine the suitable athletes for the next competitions in this sport, while giving a robust and operative overview of features for athletes selection management purposes. Finally, though our formulation, it is suitable for Tae Kwon Do athletes, but it could also be useful in other combat or martial arts.

How to cite: Lara, R. & Estévez, A. (2018). Towards an automatic detection system of sports talents: an approach to Tae Kwon Do. Sistemas & Telemática, 16(47), 31-44. https://doi.org/10.18046/syt.v16i47.32113

The authors gratefully acknowledge the contribution of the Universidad de las Fuerzas Armadas [ESPE] for the economical support for the development of this project under Research Grants 2013-PIT-014 and 2016-EXT-038.