key: cord-0072448-7zu0jk0p
authors: Echeverria, Jon; Santos, Olga C.
title: Toward Modeling Psychomotor Performance in Karate Combats Using Computer Vision Pose Estimation
date: 2021-12-15
journal: Sensors (Basel)
DOI: 10.3390/s21248378
sha: a0e8c1e2df79a665ca4b966a36c3f93b1f79509d
doc_id: 72448
cord_uid: 7zu0jk0p

Technological advances enable the design of systems that interact more closely with humans in a multitude of previously unsuspected fields. Martial arts are not outside the application of these techniques. From the point of view of the modeling of human movement in relation to the learning of complex motor skills, martial arts are of interest because they are articulated around a system of movements that are predefined, or at least, bounded, and governed by the laws of Physics. Their execution must be learned after continuous practice over time. Literature suggests that artificial intelligence algorithms, such as those used for computer vision, can model the movements performed. Thus, they can be compared with a good execution as well as analyze their temporal evolution during learning. We are exploring the application of this approach to model psychomotor performance in Karate combats (called kumites), which are characterized by the explosiveness of their movements. In addition, modeling psychomotor performance in a kumite requires the modeling of the joint interaction of two participants, while most current research efforts in human movement computing focus on the modeling of movements performed individually. Thus, in this work, we explore how to apply a pose estimation algorithm to extract the features of some predefined movements of Ippon Kihon kumite (a one-step conventional assault) and compare classification metrics with four data mining algorithms, obtaining high values with them.

Human activity recognition (HAR) techniques have proliferated and focused on recognizing, identifying and classifying inputs through sensory signals, images or video, and are used to determine the type of activity that the person being analyzed is performing [1] . Following [1] , human activities can be classified according to: (i) gestures (primitive movements of the body parts of a person that may correspond to a particular action of this person [2] ); (ii) atomic actions (movements of a person describing a certain motion that may be part of more complex activities [3] ); (iii) human-to-object or human-to-human interactions (human activities that involve two or more persons or objects [4] ); (iv) group actions (activities performed by a group of persons [5] ); (v) behaviors (physical actions that are associated with the emotions, personality, and psychological state of the individual [6] ); and (vi) events (high-level activities that describe social actions between individuals and indicate the intention or the social role of a person [7] .

HAR-type techniques are usually divided into two main groups [5, 8] : (i) HAR models based on image or video, and (ii) those that are based on signals collected from accelerometer, gyroscope, or other sensors. Due to legal and other technical issues [5] , HAR sensorbased systems are increasing their number of projects. The scope of the sensors at present is large, with cheap and good quality sensors to be able to develop any system in any sector or specific field. In this way, HAR-type systems can gather data from diverse type of sensors

The introduction of new technologies and computational approaches is providing more opportunities for human movement recognition techniques that are applied to sports. The extraction of activity data, together with their analysis with data mining algorithms is making the training efficiency higher. Through the acronym HPE (Human Pose Estimation), studies have been found in which different technologies are applied that seek to model the movement of people when performing physical activities.

Human modeling technologies have been applied for the analysis of human movements applied to sport as described in [59] . In particular, different computer vision techniques can be applied to detect athletes, estimate pose, detect movements and recognize actions [60] . In our research, we focus on movement detection. Thus, we have reviewed works that delve into existing methods: skeleton based models, contour-based models, etc. As discussed in [61] [62] [63] , 3D computer vision methodologies can be used for the estimation of the pose of an athlete in 3D, where the coordinates of the three axes (x, y, z) are necessary, and this requires the use of depth-type cameras, capable of estimating the depth of the image and video received.

Martial arts, similar to sport in general, is a discipline where human modeling techniques and HPE are being applied [64] [65] [66] . Motion capture (mocap) approaches are sometimes used and can add sensors to computer vision for motion modeling, combining the interpretation of the video image with the interpretation of the signals (accelerometer, gyroscope, magnetometer, GPS). In any case, the pose estimation work is usually performed individually [67, 68] , while the martial artist performs the techniques by themselves.

Virtual reality (VR) techniques have been used to project the avatar or the image of martial artists to a virtual environment programmed in a video game for individual practice. For instance, [69] uses computer vision techniques and transports the individual to the monitor or screen making use of background subtraction techniques [70] . In turn, [71] also creates the avatar of the person to be introduced into the virtual reality and, thus, be able to practice martial arts in front of virtual enemies that are watched through VR glasses, but in this case, body sensors are used to monitor the movements.

The application of computer vision algorithms in the martial arts domain not only focus on the identification of the pose and the movement, but there are advances in the prediction of the next attack. For this, the trainer can be monitored using residual RGB and CNN network frames to which LSTM neural networks are applied that predict the next attack movement in 2D [72] . Moreover, computer vision together with sensor data can be used to record audiovisual teaching material for Physics learning from the interaction of two (human) bodies as in Phy+Aik [73] , where Aikido techniques practiced in pairs are monitored and used to show Physics concepts of circular motion when applying a defensive technique to the attack received.

Karate [74] is a popular martial art, invited at the Tokyo Olympics, and thus, there have been efforts in applying new technologies to its modeling from a computing perspective to improve psychomotor performance. In this sense, [75] has reviewed the technologies used in twelve articles to analyze the "mawasi geri" (side kick) technique, finding that several kinds of inputs, such as 3D video image, inertial sensors (accelerometers, gyroscopes, magnetometers) and EMG sensors can be used to study the speed, position, movements of body parts, working muscles, etc. There are also studies on the "mae geri" movement (forward kick) using the Vicon optical system (with twelve MX-13 cameras) to create pattern plots and perform statistical comparison among the five expert karatekas who performed the technique [76] . Sensors are also used for the analysis of Karate movements (e.g., [77, 78] ) using Dynamic Time Warping (DTW) and Support Vector Machines (SVM). In this context, it is also relevant to know that datasets of Karate movements (such as [79] ) have been created for public use in other investigations as in the above works, but they only record individual movements.

Kinematics has also been used to analyze intra-segment coordination due to the importance of speed and precision of blows in Karate as in [80] , where a Vicon camera system consisting of seven cameras (T10 model) is used. These cameras capture the markers that the practitioners wear, and which are divided into sub-elite and elite groups. In this way, a comparison of the technical skill among both groups is made. Using a gesture description language classifier and comparing it with Markov models, different Karate techniques individually performed are analyzed to determine the precision of the model [57] .

In addition to modeling the techniques performed individually, some works have focused on the attributes needed to improve the performance in martial arts practice. For instance, VR glasses and video have been used to improve peripheral vision and anticipation [69, 81] , which has also been explored in KUMITRON [50] with computer vision filters.

The conclusion that we reached after the literature review carried out is that new technologies are being introduced in many sectors, including sports, and hence, martial arts are not an exception. The computation approaches that can be applied vary. In some cases, the signals obtained from sensors are processed, in others computer vision algorithms are used, and sometimes both are combined. In the particular case of Karate, studies are being carried out, although mainly focus on analyzing how practitioners perform the techniques individually. Thus, an opportunity has been identified to study if existing computational approaches can be applied to model the movements during the joint practice of several users, where in addition to the individual physical challenge of making the movement in the correct way, other variables such as orientation, fatigue, adaptation to the anatomy of another person, or the affective state can be of relevance. Moreover, none of the works reviewed uses the modeling of the motion to evaluate the psychomotor performance such that appropriate feedback can be provided when needed to improve the technique performed. Since computer vision seem to produce good results for HPE, in this paper we are going to explore existing computer vision algorithms that can be used to estimate the pose of karatekas in a combat, select one and use it on a dataset created with some kumite movements aimed to address the research question raised in the introduction.

In this section, we present some computer vision algorithms that can be used for pose estimation and present the methodology proposed to answer the research question.

After the analysis of the state of the art, we have identified several computer vision algorithms that produce good results for HPE, and which are listed in Table 1 . According to the literature, the top three algorithms from Table 1 offering best results are WrnchAI, OpenPose and AlphaPose. Several comparisons with varied conclusions have been made among them. According to LearnOpenCV [82] , WrnchAI and OpenPose offer similar features, although WrnchAI seems to be faster than OpenPose. In turn, [83] concludes that AlphaPose is above both when used for weight lifting in 2D vision. Other studies such as [84] find OpenPose superior and more robust when applied to real situations outside of the specific datasets such as MPII [85] and COCO [86] datasets.

From our own review, we conclude that OpenPose is the one that has generated more literature works and seems to have the largest community of developers. OpenPose has been used in multiple areas: sports [87] , telerehabilitation [88] , HAR [89] [90] [91] , artistic disciplines [92] , identification of multi-person groups [93] , and VR [94] . Thus, we have selected OpenPose algorithm for this work due to the following reasons: (i) it is open source, (ii) it can be applied in real situations with new video inputs [84] , (iii) there is a large number of projects available with code and examples, (iv) it is widely reported in scientific papers, (v) there is a strong developers community, and (vi) the API gives users the flexibility of selecting source images from camera fields, webcams, and others. It is an algorithm developed in 2018 by members of Facebook [99] that maps the pixels of the human body in 2D to turn it into a 3D surface that covers the human body. It serves one or more individuals. It is being used to determine the surface of the human body for different purposes such as trying on virtually an article of clothing on the avatar created for oneself.

(https://github.com/HRNet/ HigherHRNet-Human-Pose-Estimation, accessed on 28 November 2021)

Neural network architecture for the estimation of human pose developed in 2019 by Microsoft [100] . It is also used for semantic segmentation and object detection. Despite being a relatively new model, it is becoming a benchmark in the field of computer vision algorithms. It has been the winner in several computer vision tournaments, for example in ICCV2019 [101] . It is a useful architecture to implement in the postural analysis of televised events since it makes high-resolution estimates of postures.

OpenPose (https://github.com/CMU-Perceptual-Computing-Lab/openpose, accessed on 28 November 2021)

Computer vision algorithm for the estimation of pose in real time of several people in 2D developed in 2017 [102] . It has undergone functionalities extensions, and currently allows to be used in 3D, hand point detection, face detection, and work with Unity. The OpenPose API allows obtaining the image from various devices: recorded video, streaming video, webcam, etc. Other hardware is also supported, such as CUDA GPUs, OpenCL GPUs, and CPU-only devices.

PoseNet (https://github.com/tensorflow/ tfjs-models/tree/master/posenet, accessed on 28 November 2021)

It is a pose estimator for a single person or several people, offering 17 keypoints with which to model the human body. It was developed in 2015 [103] . At first, it was aimed at lightweight devices such as mobile phones or browsers, although today it has advanced and improved performance. WrnchAI is a human deposit estimation algorithm developed by a company based in Canada in 2014 and released only under license. It can be used for one or several individuals making use of the low latency engine, being a system compatible with all types of video. Due to its commercial use, we could not find any scientific paper describing it.

OpenPose [102] is a computer vision algorithm proposed by the Cognitive Computing Laboratory at Carnegie Mellon University for the real-time estimation of the shape of the bodies, faces and hands of various people. OpenPose provides 2D and 3D multiperson hotspot detection, as well as a calibration toolbox for estimating specific region parameters. OpenPose accepts many types of input, which can be images, videos, webcams, etc. Similarly, its output is also varied, which can be PNG, JPG, AVI or JSON, XML and YML. The input and output parameters can also be adjusted for different needs. OpenPose provides a C ++ API and works on both CPU and GPU (including versions compatible with AMD graphics cards). The main characteristics are summarized in Table 2 . In this way, we can start by exploring the 2D solutions that OpenPose offers so that it can be used with different plain cameras such as the one in a webcam, a mobile phone or even the camera of a drone (which is used in KUMITRON system [48] ). Applying 3D will require the use of depth cameras.

As described in [102] , OpenPose algorithm works as follows:

1.

Deep learning bases the estimation of pose on variations of Convolutional Neural Networks (CNN). These architectures have a strong mathematical basis on which these models are built:

Apply ReLu (REctified Linear Unit): the rectifier function is applied to increase the non-linearity in the CNN.

Group: It is based in spatial invariance, a concept in which the location of an object in an image does not affect the ability of the neural network to detect its specific characteristics. Thus, the clustering allows CNN to detect features in multiple images regardless of the lighting difference in the pictures and the different angles of the images.

Flattening: Once the grouped featured map is obtained, the next step is to flatten it. The flattening involves transforming the entire grouped feature map matrix into a single column which is then fed to the neural network for processing. 5.

Full connection: After flattening, the flattened feature map is passed through a network neuronal. This step is made up of the input layer, the fully connected layer, and the output layer. The output layer is where the predicted classes are provided. The final values produced by the neural network do not usually add up to one. However, it is important that these values are reduced to numbers between zero and one, which represent the probability of each class. This is the role of the Softmax function. 

All these steps can be represented by the following diagram in Figure 1 . : ℝ → (0,1)

for j = 1,…..,K.

(2)

All these steps can be represented by the following diagram in Figure 1 . 

The objective of the research introduced in this paper is to determine if computer vision algorithms (in this case, OpenPose) are useful to identify the movements performed in a Karate kumite by both practitioners. To answer the question, a dataset will be prepared following predefined Kihon Kumite movements both corresponding to attack and defense as explained in Section 4.2. The processing of the dataset consists of two clearly differentiated parts: (i) the training of the classification algorithms of the system using the features extracted with the OpenPose algorithm, and (ii) the application of the system to the input of non-preprocessed movements (raw data). These two stages are broken down into the following sub-stages, as shown in Figure 1 :

1. Acquisition of data input: Record the movements to create the dataset to be used in the experiment, and later to test it. 2. Feature extraction: Applying the OpenPose algorithm to the dataset to group anatomical positions of the body (called keypoints) into triplets to calculate the angle, which allow generating a pre-processed input data file for algorithm training (see Figures 3 and 4) . OpenPose allows to have the 2D position of each point (x, y). Thus, by having the coordinates of three consecutive points, the angle formed by those three with respect to the central one is calculated. An example is provided next. 3. Train a movement classifier: With the pre-processed data from point 2, train data mining algorithms to classify and identify the movements. Several data mining algo- 

The objective of the research introduced in this paper is to determine if computer vision algorithms (in this case, OpenPose) are useful to identify the movements performed in a Karate kumite by both practitioners. To answer the question, a dataset will be prepared following predefined Kihon Kumite movements both corresponding to attack and defense as explained in Section 4.2. The processing of the dataset consists of two clearly differentiated parts: (i) the training of the classification algorithms of the system using the features extracted with the OpenPose algorithm, and (ii) the application of the system to the input of non-preprocessed movements (raw data). These two stages are broken down into the following sub-stages, as shown in Figure 1 :

Acquisition of data input: Record the movements to create the dataset to be used in the experiment, and later to test it.

Feature extraction: Applying the OpenPose algorithm to the dataset to group anatomical positions of the body (called keypoints) into triplets to calculate the angle, which allow generating a pre-processed input data file for algorithm training (see Figures 3 and 4) . OpenPose allows to have the 2D position of each point (x, y). Thus, by having the coordinates of three consecutive points, the angle formed by those three with respect to the central one is calculated. An example is provided next.

Train a movement classifier: With the pre-processed data from point 2, train data mining algorithms to classify and identify the movements. Several data mining algorithms can be used for the classification in the current work generating the corresponding learning models (see below). For the evaluation of the classification performance, 10-fold cross validation is proposed, following [105] [106] [107] .

Test the movement classifier: Apply the trained classifier to the non-preprocessed input (raw data) with the movements performed by the karateka. This process follows the scheme in Figure 2 , which shows the steps proposed for Karate movement recognition using OpenPose algorithm to extract the features from the images and data mining algorithms on these features to classify the movements: rithms can be used for the classification in the current work generating the corresponding learning models (see below). For the evaluation of the classification performance, 10-fold cross validation is proposed, following [105] [106] [107] This process follows the scheme in Figure 2 , which shows the steps proposed for Karate movement recognition using OpenPose algorithm to extract the features from the images and data mining algorithms on these features to classify the movements: OpenPose works with 25 keypoints (numbered from 0 to 24, as shown in Section 4.3 in Figure 9 ) for the extraction of input data, which are used to generate triples, as it can be seen in Figure 3 . OpenPose keypoints are grouped into triplets of three consecutive points, which are the input attributes (angles) for classifying the output classes. To define the keypoints, OpenPose expanded the 18 keypoints of COCO dataset (https://cocodataset.org/#home accessed on 2 December 2021) with the ones for the feet and waist from the Human Foot Keypoint dataset (https://cmu-perceptual-computinglab.github.io/foot_keypoint_dataset/ accessed on 2 December 2021)). OpenPose works with 25 keypoints (numbered from 0 to 24, as shown in Section 4.3 in Figure 9 ) for the extraction of input data, which are used to generate triples, as it can be seen in Figure 3 . OpenPose keypoints are grouped into triplets of three consecutive points, which are the input attributes (angles) for classifying the output classes. To define the keypoints, OpenPose expanded the 18 keypoints of COCO dataset (https://cocodataset.org/#home accessed on 2 December 2021) with the ones for the feet and waist from the Human Foot Keypoint dataset (https://cmu-perceptual-computing-lab.github.io/foot_keypoint_ dataset/ accessed on 2 December 2021)). Sensors 2021, 21, x FOR PEER REVIEW 10 of 28 This allows identifying the numbering of the triples with their body position. As an example we have the keypoints numbered 5, 6 and 7 that correspond to the main part "Shoulders Left Arm" as it can be seen in Figure 4 . In Figure 3 , the angular creation of the commented triplet can be graphically observed. Figure 4 collects all the triples formed by the keypoints in Figure 3 , naming them the main body part they represent. The grouping of the keypoints in triples makes it possible to avoid the variability of the points depending on the height of the practitioners, in order to calculate their angle. In this way, a person This allows identifying the numbering of the triples with their body position. As an example we have the keypoints numbered 5, 6 and 7 that correspond to the main part "Shoulders Left Arm" as it can be seen in Figure 4 . In Figure 3 , the angular creation of the commented triplet can be graphically observed. Figure 4 collects all the triples formed by the keypoints in Figure 3 , naming them the main body part they represent. The grouping of the keypoints in triples makes it possible to avoid the variability of the points depending on the height of the practitioners, in order to calculate their angle. In this way, a person This allows identifying the numbering of the triples with their body position. As an example we have the keypoints numbered 5, 6 and 7 that correspond to the main part "Shoulders Left Arm" as it can be seen in Figure 4 . In Figure 3 , the angular creation of the commented triplet can be graphically observed. Figure 4 collects all the triples formed by the keypoints in Figure 3 , naming them the main body part they represent. The grouping of the keypoints in triples makes it possible to avoid the variability of the points depending on the height of the practitioners, in order to calculate their angle. In this way, a person of 190 cm tall and another of 170 cm tall, will have similar angles for the same Karate posture, compared to the bigger variation of the keypoints due to the different height of the bodies. This allows the algorithm training to be more efficient and fewer inputs are required to achieve optimal computational performance.

As a result, Figure 5 shows the processing flow proposed to answer the research question regarding the identification of kumite postures. On the top part of the figure, the data models are trained to identify all the movements selected for the study. This is done by generating a dataset with the angles of each Karate technique, which is trained according to different data mining algorithms. Subsequently, these trained algorithms are used on real kumite movements in real time in order to identify the kumite posture performed, as shown on the bottom. For the selection of the data mining classifiers, a search was made of the types of classifiers applied in general to computer vision. Some works such as [108] [109] [110] [111] [112] use machine learning (ML) algorithms, mainly decision trees and bayesian networks. However, deep learning (DL) techniques are more and more used for the identification and qualification of video images as in [113] [114] [115] [116] . In particular, a deep learning classification algorithm that is having very good results according to the studies found is the Weka DeepLearning4j algorithm [117] [118] [119] . The Weka Deep Learning kit allows to create advanced Neural Network Layers as it is explained in [120] .

In addition to a general analysis of ML and DL algorithms, specific application to sports, human modeling and video image processing were also sought. The BayesNet algorithm has already been used to estimate human pose in other similar experiments [121] [122] [123] . Another algorithm that has been applied to this type of classifications is the J48 decision tree [124] [125] [126] . These two ML algorithms were selected for the classification of the movements, as well as two neural network algorithms included in the WEKA application: the MultiLayerPerceptron (MLP) algorithm, which has also been used in different works as well as similar algorithms [127, 128] , and the aforementioned DeepLearning4J. Thus, we have selected two ML and two DL algorithms. Table 3 compares both types of algorithms following [129] . Due to the type of hardware that was used, which consisted of standard cameras (without depth functionality) available in webcams, mobiles and drones, the research was oriented to the application of the video image in 2D format for its correct processing and postural identification.

The operation of OpenPose for the application programmed consists of a communication system between OpenPose and the functions designed for the video manipulation, done with the four classes in a Java application listed below. The computational cost associated with the processing is determined through the computational analysis of the following main parts of the code that are represented in Figure 6 . as well as similar algorithms ( [127, 128] ), and the aforementioned DeepLearning4J. Thus, we have selected two ML and two DL algorithms. Table 3 compares both types of algorithms following [129] . Due to the type of hardware that was used, which consisted of standard cameras (without depth functionality) available in webcams, mobiles and drones, the research was oriented to the application of the video image in 2D format for its correct processing and postural identification.

The operation of OpenPose for the application programmed consists of a communication system between OpenPose and the functions designed for the video manipulation, done with the four classes in a Java application listed below. The computational cost associated with the processing is determined through the computational analysis of the following main parts of the code that are represented in Figure 6 . The video input is captured by the OpenPose application libraries that identify the keypoints of each individual that appears in the image. As explained in the official Open-Pose paper [102] , the runtime of OpenPose consists of two major parts: (1) CNN processing time whose complexity is O(1), constant with varying number of people; and (2) multi-person parsing time, whose complexity is O(N 2 ), where N represents the number of people.

OpenPose communicates with our application by sending information about the identified keypoints and the number of individuals in the image. The Java application, receives the data from OpenPose, dimensioned in "N" the problem for the complexity calculation.

There are four main classes: The video input is captured by the OpenPose application libraries that identify the keypoints of each individual that appears in the image. As explained in the official OpenPose paper [102] , the runtime of OpenPose consists of two major parts: (1) CNN processing time whose complexity is O(1), constant with varying number of people; and (2) multi-person parsing time, whose complexity is O(N 2 ), where N represents the number of people.

OpenPose communicates with our application by sending information about the identified keypoints and the number of individuals in the image. The Java application, receives the data from OpenPose, dimensioned in "N" the problem for the complexity calculation.

There are four main classes: • Full cost per frame for more than one person: O(Nˆ2) + 2 * O(ni) + K; It can be seen how in this case the higher computational cost of the OpenPose algorithm marks the total operational cost of the set of applications, finally reaching a quadratic order.

To answer the research question, first we have to select the video images to generate the dataset with the interaction of the karatekas. There are different forms of kumite in Karate, from multi-step combat (to practice) to free kumite [130] (very explosive with very fast movements [131] ). Multi-step combat (each step is as an attack) is a simple couple exercise that slowly lead the karateka to an increasingly free action, according to predefined rules. It does not necessarily have to be equated with competition: it is more of a pair exercise in which the participant together with a partner develops a better understanding of the psychomotor technique. Participants do not compete with each other but train with themselves. The different types of kumite are listed in Table 4 , from those with less freedom of movement to free combat. In order to follow a progressive approach in our research, we started with the Ippon Kihon kumite, which is a pre-established exercise so that it can facilitate the labeling of movements and their analysis. This is the most basic kumite exercise, consisting of the conventional one-step assault. We will use it to compare the results obtained from the application of the data mining algorithms on the feature extracted from the videos recorded and processed using the OpenPose algorithm.

In this first step of our research, the following attack and defense sequences were defined to create the initial dataset, as shown in Figure 7 .

The Ippon Kihon kumite sequence proposed for this study would actually be (i) Gedan Barai, (ii) Oi Tsuki, (iii) Soto Uke, and (iv) Gyaku Tsuki. However, to calibrate the algorithm and enrich the dataset, we also took two postures Kamae (which is the starting posture), both for attack and defense, although they are not properly part of the Ippon Kihon kumite.

In this first step of our research, the following attack and defense sequences were defined to create the initial dataset, as shown in Figure 7 . The Ippon Kihon kumite sequence proposed for this study would actually be (i) Gedan Barai, (ii) Oi Tsuki, (iii) Soto Uke, and (iv) Gyaku Tsuki. However, to calibrate the algorithm and enrich the dataset, we also took two postures Kamae (which is the starting posture), both for attack and defense, although they are not properly part of the Ippon Kihon kumite.

Since monitoring the simultaneous interaction of two karatekas in movement is complex, we have made this first dataset as simple as possible to familiarize ourselves with the algorithm, learn and understand the strengths and weaknesses of OpenPose and explore its potential in the human movement computing scenario addressed in this research (performing martial arts combat techniques between two practitioners). We aim at increasing the difficulty and complexity of the dataset gradually along the research. Thus, for this first dataset, we selected direct upper trunk techniques (not lateral or angular) to facilitate the work of OpenPose classification. In addition, simple techniques with the limbs were selected. Thus, when working in 2D, lateral and circular blows that could acquire angles that could be difficult to calculate in their trajectory and execution, are avoided. The sequence of movements in pairs of attack and defense is shown in Table 5 . 

Once the dataset was prepared, a Java application was developed in KUMITRON to collect the data from the practitioners. In particular, in this initial and exploratory collection of data, a green belt participant was video recorded performing the proposed movements, both attack and defense. The video was taken statically for one minute in which the karateka had to be in one for the predefined postures of Table 5 and Figure 7 , and the camera was moved to capture different possible angles of the shot in 2D. In addition, videos were also taken dynamically in which the karateka went from one posture to another and waited 30 seconds in the last one. Since monitoring the simultaneous interaction of two karatekas in movement is complex, we have made this first dataset as simple as possible to familiarize ourselves with the algorithm, learn and understand the strengths and weaknesses of OpenPose and explore its potential in the human movement computing scenario addressed in this research (performing martial arts combat techniques between two practitioners). We aim at increasing the difficulty and complexity of the dataset gradually along the research. Thus, for this first dataset, we selected direct upper trunk techniques (not lateral or angular) to facilitate the work of OpenPose classification. In addition, simple techniques with the limbs were selected. Thus, when working in 2D, lateral and circular blows that could acquire angles that could be difficult to calculate in their trajectory and execution, are avoided. The sequence of movements in pairs of attack and defense is shown in Table 5 . 

Once the dataset was prepared, a Java application was developed in KUMITRON to collect the data from the practitioners. In particular, in this initial and exploratory collection of data, a green belt participant was video recorded performing the proposed movements, both attack and defense. The video was taken statically for one minute in which the karateka had to be in one for the predefined postures of Table 5 and Figure 7 , and the camera was moved to capture different possible angles of the shot in 2D. In addition, videos were also taken dynamically in which the karateka went from one posture to another and waited 30 seconds in the last one. Figure 8 shows the interface of the application while recording the Oi Tsuki movement performed by the karateka while testing the application. On the left panel, the postures detected are listed chronologically from bottom to top. On the right of the image, the skeleton identified with the OpenPose algorithm is shown on the real image. Figure 8 shows the interface of the application while recording the Oi Tsuki movement performed by the karateka while testing the application. On the left panel, the postures detected are listed chronologically from bottom to top. On the right of the image, the skeleton identified with the OpenPose algorithm is shown on the real image. The application can also identify the skeletons of both karatekas when interacting in an Ippon Kihon kumite. Figure 9 shows the left practitioner (male, green belt) launching a Gedan Barai attack when the one on the right (women, white belt) is waiting in Kamae. 

The attributes selected for the classification with the data mining algorithms, as explained in Section 3.2, were the 26 triples obtained from the OpenPose keypoints. In this way, it is possible to calculate the angles that are formed by the different joints of the body of each karateka. Figure 10 shows the 26 different attributes considered corresponding to the 26 triplets introduced in Figure 4 . They are shown in order as they are sent by the application. The last window shows the classes (postures) that are being identified. Colors are assigned as indicated in Table 6 . The application can also identify the skeletons of both karatekas when interacting in an Ippon Kihon kumite. Figure 9 shows the left practitioner (male, green belt) launching a Gedan Barai attack when the one on the right (women, white belt) is waiting in Kamae. The application can also identify the skeletons of both karatekas when interacting in an Ippon Kihon kumite. Figure 9 shows the left practitioner (male, green belt) launching a Gedan Barai attack when the one on the right (women, white belt) is waiting in Kamae. 

The attributes selected for the classification with the data mining algorithms, as explained in Section 3.2, were the 26 triples obtained from the OpenPose keypoints. In this way, it is possible to calculate the angles that are formed by the different joints of the body of each karateka. Figure 10 shows the 26 different attributes considered corresponding to the 26 triplets introduced in Figure 4 . They are shown in order as they are sent by the application. The last window shows the classes (postures) that are being identified. Colors are assigned as indicated in Table 6 . 

The attributes selected for the classification with the data mining algorithms, as explained in Section 3.2, were the 26 triples obtained from the OpenPose keypoints. In this way, it is possible to calculate the angles that are formed by the different joints of the body of each karateka. Figure 10 shows the 26 different attributes considered corresponding to the 26 triplets introduced in Figure 4 . They are shown in order as they are sent by the application. The last window shows the classes (postures) that are being identified. Colors are assigned as indicated in Table 6 . Table 6 . Defined Karate postures with the colors used in Figure 10 and the number of dataset inputs considered.

Color Number of Dataset Inputs Attack01: "Kamae"

Attack02: "Gedan Barai"

Attack03: "Oi Tsuki"

Defense01: "Kamae"

Defense02: "Soto Uke"

Defense03: "Gyaku Tsuki"

For each of the postures (classes) several videos were recorded with KUMITRON to create the initial dataset (as described in Section 4.3) and obtain the video inputs of the postures for the OpenPose algorithm indicated in Table 6 . The postures were recorded in different ways. On the one hand, one minute long videos were made where the camera moved from one side to the other while the karateka was completely still maintaining the posture. To generate postural variability, other videos were created where the karateka performed the selected postures for the Ippon Kihon kumite, stopping at the one that was For each of the postures (classes) several videos were recorded with KUMITRON to create the initial dataset (as described in Section 4.3) and obtain the video inputs of the postures for the OpenPose algorithm indicated in Table 6 . The postures were recorded in different ways. On the one hand, one minute long videos were made where the camera moved from one side to the other while the karateka was completely still maintaining the posture. To generate postural variability, other videos were created where the karateka performed the selected postures for the Ippon Kihon kumite, stopping at the one that was to be recorded. In this way, it is assumed that the recorded videos can offer more difference between keypoints than just making static recordings. From the recorded videos, the inputs generated come out at a rate of 25 frames per second. Thus, from one second of video, 25 inputs are generated for every second. A feature in OpenPose to clean the inputs with values very different from the average ones was used to deal with the members that remain in the back of the camera.

In Figure 10 , it can be observed in a range of 180 degrees (range in which the 26 angles move) the values that each class takes according to the part of the body. This can be interesting in future research to investigate, for example, whether a posture is overloaded and is likely to cause some type of injury because it is generating an angle that what it is convenient for that part of the body.

The statistical results obtained by applying the four algorithms (BayesNet, J48, MLP and DeepLearning4J) using Weka suite to the identification of movements of the Ippon Kihon kumite dataset built in this first step of our research are presented in Table 7 . The dataset obtained for the expression is publicly accessible (https://github.com/Kumitron accessed on 2 December 2021). Table 7 shows that the classifying algorithms have obtained a good classification performance without significant differences among them. Nonetheless, the performance of the algorithms will need to be revaluated when more users and more movements are included in the dataset and it becomes more complex. The detail of the evaluation metrics results by type of algorithm is shown in Table 8 . As expected from Table 7 , evaluation metrics are good in all the algorithms. DeepLearn-ing4J seems to perform slightly better. However, another important variable to consider in the performance analysis between algorithms is the time it takes to build the models. Figure 11 shows that, from the comparison of the four algorithms, the differences in processing time between the ML and DL algorithms are high. As expected, neural networks take much longer, and since performance results are similar, they do not seem necessary for the current analysis. However, we will analyze how both the processing time evolves in these algorithms, with a greater number of inputs and varying conditions of speed of movements in future versions of the dataset. The detailed performance results by type of movement (class) for each algorithm is shown in Tables 9−12 reporting the values of the metrics computed by Weka. These results were already good in the overall analysis, and no significant difference was found between types of movements in any of the algorithms. The metrics computed for the BayesNet algorithm are reported in Table 9 . The metrics computed for the J48 algorithm are reported in Table 10 . Figure 11 . Comparative times between the applied algorithms to build the models.

The detailed performance results by type of movement (class) for each algorithm is shown in Tables 9-12 reporting the values of the metrics computed by Weka. These results were already good in the overall analysis, and no significant difference was found between types of movements in any of the algorithms.

The metrics computed for the BayesNet algorithm are reported in Table 9 . The metrics computed for the J48 algorithm are reported in Table 10 . The metrics computed for the MLP algorithm are reported in Table 11 . The metrics computed for the BayesNet algorithm are reported in Table 12 . 

For the application of the classification models, possible methods of optimization of the hyperparameters were investigated [132] . Since the results obtained with the current dataset and default parameters in Weka were good, no optimization was performed, and default values were left. Table 13 shows the scripts used to call the algorithms in Weka. 

Recalling the classification of human activities introduced in Section 1 [1] , this reesearch focuses on type iv, that is, group activities, which focus on the interaction of actions between two people. That is the line of research that makes this work different from those found in the state of the art. The application of OpenPose algorithm to kumite recorded videos aims to facilitate improving the practitioners' technique against different opponents when applying the movements freely and in real time. This is important for the assimilation of the techniques in all kinds of scenarios, including sport practice and personal defense situations. The modeling of movements for psychomotor performance can provide a useful tool to learn Karate that opens up new lines of research.

To start with, it can allow studying and improving the combat strategy, having exact measurements of the distance necessary to know when a fighter can be hit by the opponent's blows, as well as the distance necessary to reach the opponent with a blow. This allows training in the gym in a scientific and precise way for combat preparation and distance taking on the mat. In addition, studies can be carried out on how some factors such as fatigue can create variability in the movements developed within a combat in each one of its phases. This is important because this allows for deciding on the choice of certain fighting techniques at the beginning or end of the kumite to win. Modeling and capturing tagged movements can also be used to produce datasets and to apply them in other areas such as cinema or video games, which are economically attractive sectors.

The current approach to progress in our research is to label not only single postures (techniques) within the Ippon Kihon kumite (as done here) but the complete sequences of the movements. There is some work that can guide the technical implementation of this approahc, where classes are first identified and then become subclasses of a superclass [133] . In the work reported in this paper, classes have been defined as the specific techniques to be identified. In the next step, the idea is that these techniques are considered as subclasses, being part of a superior class or superclass. In this way, when the system identifies a concatenation of specific movements, it should be able to classify the attack/defense sequence (default Ippon Kihon kumite) that is being carried out as shown in Figure 12 .

Recalling the classification of human activities introduced in Section 1 [1] , this reesearch focuses on type iv, that is, group activities, which focus on the interaction of actions between two people. That is the line of research that makes this work different from those found in the state of the art. The application of OpenPose algorithm to kumite recorded videos aims to facilitate improving the practitioners' technique against different opponents when applying the movements freely and in real time. This is important for the assimilation of the techniques in all kinds of scenarios, including sport practice and personal defense situations. The modeling of movements for psychomotor performance can provide a useful tool to learn Karate that opens up new lines of research.

To start with, it can allow studying and improving the combat strategy, having exact measurements of the distance necessary to know when a fighter can be hit by the opponent's blows, as well as the distance necessary to reach the opponent with a blow. This allows training in the gym in a scientific and precise way for combat preparation and distance taking on the mat. In addition, studies can be carried out on how some factors such as fatigue can create variability in the movements developed within a combat in each one of its phases. This is important because this allows for deciding on the choice of certain fighting techniques at the beginning or end of the kumite to win. Modeling and capturing tagged movements can also be used to produce datasets and to apply them in other areas such as cinema or video games, which are economically attractive sectors.

The current approach to progress in our research is to label not only single postures (techniques) within the Ippon Kihon kumite (as done here) but the complete sequences of the movements. There is some work that can guide the technical implementation of this approahc, where classes are first identified and then become subclasses of a superclass [133] . In the work reported in this paper, classes have been defined as the specific techniques to be identified. In the next step, the idea is that these techniques are considered as subclasses, being part of a superior class or superclass. In this way, when the system identifies a concatenation of specific movements, it should be able to classify the attack/defense sequence (default Ippon Kihon kumite) that is being carried out as shown in Figure 12 . This can be useful in order to add personalized support to the kumite training based on the analysis of the psychomotor performance both comparing the current movements with a good execution and analyzing the temporal evolution of the execution of the movements. The system will be able to recognize if the karateka is developing the movements This can be useful in order to add personalized support to the kumite training based on the analysis of the psychomotor performance both comparing the current movements with a good execution and analyzing the temporal evolution of the execution of the movements. The system will be able to recognize if the karateka is developing the movements of a certain section correctly, not only one technique, but the entire series. Thus, training is expected to help practitioners assimilate the concatenations of movements in attack and defense. Furthermore, having super classes identified by the system makes it possible to know in advance which next subclass will be carried out, in such a way that the following movements that will be carried out by a karateka can be monitored through the system, provided that they follow the pattern of some predefined concatenated movements. This is of special interest to be able to work the anticipation training for the defender.

For this, a methodology has to be followed for creating super classes that bring together the different concatenated movements. In principle, it seems to be technically possible, as it has been introduced above, but more work is necessary to be able to answer the question with absolute security, since in order to recognize a superclass, the application must recognize all subclasses without any errors or identification of any wrong postures in the sequence. This requires high identification and classification precision. Moreover, the system needs to have in memory the different movements of each defined superclass, which means that it knows which movement should be the next to perform. In this way, training can be enriched with attack anticipation work in a dynamic and natural way.

For future work, and in addition to exploring other classification algorithms, such as the LSTM algorithm that have reported good results in other works [47, 134] , it would be interesting to add more combinations of movements to create a wider base of movements and to compare the results between different Ippon Kihon kumite movements. Thus, the importance of the different keypoints from OpenPose will be studied (including the analysis of the movements that are more difficult to identify) as well as if it is possible to eliminate some to make the application lighter. For the Ippon Kihon kumite sequence of postures that has been used in the current work, the keypoints of the legs does not seem to be especially important since the labeled movements are few, and the monitored postures are well identified only with the upper part of the body.

Next, the idea is to extend the current research to faster movements that are performed in other types of kumite, following the order of difficulty of these, from Kihon Sambon kumite to competition kumite, or kumite with free movements. This will allow the calibration of the technical characteristics of the algorithm used (currently OpenPose) to check what type of image speed is capable of working with while still obtaining satisfactory results. This would also be an important step forward in adding elements to anticipation and peripheral vision training. Such attributes have already been started to be explored in our research using OpenCV algorithms [50] .

The progress in the classification of the movements during a kumite will be integrated into KUMITRON intelligent psychomotor system [48] . KUMITRON collects both video and sensor data from karatekas' practice in a combat, models the movement information, and after designing the different types of feedback that can be required (e.g., with TORMES methodology [135] ), delivers the appropriate one for each karateka in each situation (e.g., taking into account the karatekas' affective state during the practice, which can be obtained with the physiological sensors available in KUMITRON following a similar approach as in [136] ) through the appropriate sensorial channel (vibrotactile feedback should be explored due to the potential discussed in [137] ). In this way, it is expected that computer vision support in KUMITRON can help karatekas learn how to perform the techniques in a kumite with the explosiveness required to win the point, making rapid and strong movements that quickly react in real time and in a fluid way to the opponent's technique, adapting also the movement to the opponent's anatomy. The psychomotor performance of the karatekas is to be evaluated both comparing the current movements with a good execution and analyzing the temporal evolution of the execution of the movements.

The main objective of this work was to carry out a first step in our research to assess if computer vision algorithms allow identifying the postures performed by karatekas within the explosive movements that are developed during a kumite. The selection of kumite was not accidental. It was chosen because it challenges image processing in human movement computing, and there is little scientific literature on modeling the psychomotor performance in activities that involve the joint participation of several individuals, as in a karate combat. In addition, the movements performed by a karateka in front of an opponent may vary with respect to performing them alone through katas due to factors such as fear, concentration or adaptation to the opponent's physique (e.g., height).

The results obtained from the training of the classification algorithms with the features extracted from the recorded videos of different Kihon kumite postures and their application to non labelled images have been satisfactory. It has been observed that the four algorithms used to classify the features extracted with OpenPose algorithm for the detection of movements (i.e., BayesNet, J48, MLP and DeepLearning4j in Weka) have a precision of above 99% for the current (and limited) dataset. With this percentage of success, it is expected that progress can be made with the inclusion of more Kihon kumite sequences in a new version of the dataset to analyze whether they can be identified within a Kihon kumite fighting exercise. In this way, the research question would be satisfied for the training level of Kihon kumite.

A review of human activity recognition methods

Discovering motion primitives for unsupervised grouping and one-shot learning of human actions, gestures, and expressions

Motion Part Regularization: Improving Action Recognition via Trajectory Group Selection

Structured learning of human interactions in TV shows

Segmentation and Recognition of Basic and Transitional Activities for Continuous Physical Human Activity

Don't classify ratings of affect

Social roles in hierarchical models for human activity recognition

Deep learning algorithms for human activity recognition using mobile and wearable sensor networks: State of the art and research challenges

A new approach to human activity recognition using machine learning techniques

Wearable Computing: Accelerometers' Data Classification of Body Postures and Movements

Human Activity Recognition Using Smartphone Sensors: A Dense Neural Network Approach

A robust feature extraction model for human activity characterization using 3-axis accelerometer and gyroscope data

Phased Human Activity Recognition based on GPS

Improved use of foot force sensors and mobile phone GPS for mobility activity recognition

Pulse-wave monitoring by means of focused laser beams scattered by skin surface and membranes. Static Dyn. Light Scatt

Characterizing human activity induced impulse and slip-pulse excitations through structural vibration

A preliminary study of sensing appliance usage for human activity recognition using mobile magnetometer

Human activity recognition using inertial/magnetic sensor units

Body temperature-Indoor condition monitor and activity recognition by mems accelerometer based on IoT-alert system for people in quarantine due to COVID-19

Artificial Intelligence in Psychomotor Learning: Modeling Human Motion from Inertial Sensor Data

Automated eloquent cortex localization in brain tumor patients using multi-task graph neural networks

Human Motion Analysis: A Review

Human activity analysis: A review

Sensors, vision and networks: From video surveillance to activity recognition and health monitoring

Multimodal human activity recognition for industrial manufacturing processes in robotic workcells

Human Activity Recognition for Domestic Robots

Human activity recognition in pervasive health-care: Supporting efficient remote collaboration

IoT based mobile healthcare system for human activity recognition

A survey on wearable sensor modality centred human activity recognition in health care

A framework for the integration of gesture and posture recognition using HMM and SVM

Artificial Intelligence of Things in Sports Science: Weight Training as an Example

Collective Sports: A multi-task dataset for collective activity recognition

Sensor fusion for multi-player activity recognition in game sports

Use of wearable technologies for analysis of activity recognition for sports

Trends Supporting the In-Field Use of Wearable Inertial Sensors for Sport Performance Evaluation: A Systematic Review

Racquet sports recognition using a hybrid clustering model learned from integrated wearable sensor

Naive Bayes approach to predict the winner of an ODI cricket game

Temporal post-processing of decision tree outputs for sports video categorisation

Event detection in field sports video using audio-visual features and a support vector machine

Light sport exercise detection based on smartwatch and smartphone using k-Nearest Neighbor and Dynamic Time Warping algorithm

Activity Recognition in Surfing-A Comparative Study between Hidden Markov Model and Support Vector Machine

Shot classification of field sports videos using AlexNet Convolutional Neural Network

Intelligent tutoring systems for psychomotor training-A systematic literature review

Intelligent systems for psychomotor learning. In Handbook of Artificial Intelligence in Education

Training the Body: The Potential of AIED to Support Personalized Motor Skills Learning

The Full Life Cycle of Adaptation in aLFanet eLearning Environment

KSAS: A Mobile App with Neural Networks to Guide the Learning of Motor Skills

Artificial intelligence system to monitor karate fights that synchronize aerial images with physiological and inertial signals

A Multimodal Psychomotor Intelligent Learning System to Provide Personalized Support when Training Karate Combats. MAIEd'21 Workshop. The First International Workshop on Multimodal Artificial Intelligence in Education

Punch Anticipation in a Karate Combat with Computer Vision

Learning in Martial Arts: An opportunity for User Modeling, Adaptation and Personalization

Can Aikido Help with the Comprehension of Physics? A First Step towards the Design of Intelligent Psychomotor Systems for STEAM Kinesthetic Learning Scenarios

My Way of Life

World Karate Federation. Karate Competition Rules

Application of Hidden Markov Models and Gesture Description Language classifiers to Oyama karate techniques recognition

Relationship Between Functional Movement Screen and Some Athletic Abilities in Karate Athletes

Application of assistive computer vision methods to Oyama karate techniques recognition

Beyond Cognitive and Affective Issues: Designing Smart Learning Environments for Psychomotor Personalized Learning

Fast human pose estimation

Recent advances in computer vision

Monocular 3D Human Motion Capture and Fine-Grained Understanding in Challenging Sports Videos

Animation of 3D Human Model Using Markerless Motion Capture Applied To Sports

Multi-person 3D pose estimation and tracking in sports

Mocap data editing via movement notations

An Evaluation of Pose Estimation in Video of Traditional Martial Arts Presentation

A Review of Virtual Reality and Motion Capture in Martial Arts Training

Martial Arts, Dancing and Sports dataset: A challenging stereo and multi-view dataset for 3D human pose estimation

Biomechanics Analysis of Combat Sport (Silat) by Using Motion Capture System

Training using virtual reality improves response behavior in karate kumite

Wallflower: Principles and practice of background maintenance

Martial Arts Training in Virtual Reality with Full-body Tracking and Physically Simulated Opponents

Martial arts in artificial reality

Intelligent Framework for Learning Physics with Aikido (Martial Art) and Registered Sensors

Available online: www.kenkyoha.com (accessed on

Biomechanical Analysis of Mawashi-Geri in Technique in Karate: Review Article

Visualized movement patterns and their analysis to classify similaritiesdemonstrated by the karate kick Mae-geri

Human actions analysis: Templates generation, matching and visualization applied to motion capture of highly-skilled karate athletes

Recognition punches in karate using acceleration sensors and convolution neural networks

A multimodal dataset for the analysis of movement qualities in karate martial art

Coordinative intra-segment indicators of karate performance

Improvement of Early Recognition of Attacks in Karate Kumite Due to Training in Virtual Reality

Pose Detection Comparison: WrnchAI vs OpenPose

Human Pose Estimation Assisted Fitness Technique Evaluation System

Filling the gaps: Predicting missing joints of human poses using denoising autoencoders

2D Human Pose Estimation: New Benchmark and State of the Art Analysis

Microsoft coco: Common objects in context

Imagery based Parametric Classification of Correct and Incorrect Motion for Push-up Counter Using OpenPose

Applying Vision-Based Pose Estimation in a Telerehabilitation Application

Fall Detection Based on Key Points of Human-Skeleton Using OpenPose

Human Motion Prediction by 2D Human Pose Estimation using OpenPose

A framework for fall detection based on OpenPose skeleton and LSTM/GRU models

Application Development of Dance Pose Recognition Based on Embedded Artificial Intelligence Equipment

Multi person pose estimation based on improved openpose model

Mobile human shape superimposi-tion: An initial approach using OpenPose

RMPE: Regional Multi-Person Pose Estimation

Joint subset partition and labeling for multi person pose estimation

Scalable object detection using deep neural networks

Human Pose Estimation via Deep Neural Networks

DensePose: Dense Human Pose Estimation in the Wild

Deep High-Resolution Representation Learning for Visual Recognition

Joint COCO and Mapillary Workshop at ICCV 2019: COCO Keypoint Detection Challenge Track

Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields

PoseNet: A convolutional network for real-time 6-dof camera relocalization

Computer Vision-Driven Evaluation System for Assisted Decision-Making in Sports Training Lijin

Cross-validation

Sensitivity Analysis of k-Fold Cross Validation in Prediction Error Estimation

Reliable Accuracy Estimates from k-Fold Cross Validation

News video classification using multimodal classifiers and text-biased combination strategies

Adapting SVM classifiers to data with shifted distributions

Tree-based classifiers for bilayer video segmentation

Who are you?-Learning person specific classifiers from video

Computer vision based local fruit recognition

Everything You Wanted to Know about Deep Learning for Computer Vision but Were Afraid to Ask

Deep Learning for Computer Vision: A Brief Review

Application of deep learning to computer vision: A comprehensive study

The application of deep learning in computer vision

A spatially explicit deep learning neural network model for the prediction of landslide susceptibility

Music emotion recognition using recurrent neural networks and pretrained models

White learning methodology: A case study of cancer-related disease factors analysis in real-time PACS environment

High performance machine learning and data science based implementation using Weka. Mater

Bayesian capsule networks for 3D human pose estimation from single 2D images

A Two-Stage Bayesian Network Method for 3D Human Pose Estimation from Monocular Image Sequences

A Non-parametric Bayesian Network Prior of Human Pose

Identifying lameness in horses through deep learning

A Kidnapping Detection Using Human Pose Estimation in Intelligent Video Surveillance Systems

Real-Time Human Pose and Gesture Recognition for Autonomous Robots Using a Single Structured Light 3D-Scanner

Deep neural networks for human pose estimation from a very low resolution depth image

3D human pose estimation using convolutional neural networks with 2D pose information

A Survey of Video Based Action Recognition in Sports

Karate Fighting Techniques: The Complete Kumite

Physical and Physiological Profile of Elite Karate Athletes

Automatic Model Selection and Hyperparameter Optimization in WEKA

Human Activity Recognition in Untrimmed Video using Deep Learning for Sports Domain

Bayesian graph convolution LSTM for skeleton based action recognition

Practical guidelines for designing and evaluating educationally oriented recommendations

An Open Sensing and Acting Platform for Context-Aware Affective Support in Ambient Intelligent Educational Settings

Toward personalized vibrotactile support when learning motor skills

The authors would like to thank the Karate practitioners who have given feedback along the design and development of KUMITRON, especially Victor Sierra and Marcos Martín at IDavinci, who have been following this research from the beginning.

The authors declare no conflict of interest.