Commonality of Motions Having Effective Features with

Respect to Methods for Identifying the Moves Made during

Kumite Sparring in Karate

Keiichi Sato

, Hitoshi Matsubara

and Keiji Suzuki

National Institute of Technology, Hakodate College, Japan

Graduate School of Information Science and Technology, The University of Tokyo, Japan

Faculty of Systems Information Science, Future University Hakodate, Japan

Keywords: Image Recognition, Deep Learning, Convolutional Neural Network.

Abstract: In recent years, active use of imaging technology, sensor technology, and AI (artificial intelligence) has been

on the rise to identify various plays that occur in sports competition to assist judges, coaches, and athletes.

However, no study result has been reported on this research topic as it pertains to karate sparring competition.

The lack of past studies on this topic is attributable to the fact that no viable method has been developed to

acquire motion data or to identify athletes’ motions at a fundamental level in competition, as no sensors can

be worn by athletes on their bodies, and there are a number of blind spots that occur in competition due to

fast-paced exchanges that competing athletes engage in, among other factors. Therefore, in this study, footage

of kumite (sparring) in the practice of karate simulating actual competitive matches was captured using video

cameras to conduct a motion identification experiment using a CNN (convolutional neural network). After

comparing the result of this study to that of previous motion identification experiments in which subjects wore

sensors on their bodies, it has been determined that the motions having effective features are common between

the two types of experiments.

1 INTRODUCTION

No published studies have been conducted on the

kumite competition of karate. One of the main reasons

is because no viable method to accurately sense and

identify the various motions that occur in karate

sparring contests has been developed to date. Another

reason is because a method is yet to be invented that

enables acquisition of appropriate motion data for

motion identification in a manner that does not

interfere with the contests that are taking place. These

are the two main objects. Therefore, in this study, a

video-camera-based motion data acquisition method

is suggested that won’t affect the karate sparring

matches themselves, and also a deep-learning-

powered method that will form a basis for identifying

the various motions of karate sparring, to achieve the

aforementioned two objectives. In related studies,

https://orcid.org/0000-0002-6712-4505

https://orcid.org/0000-0002-3104-3025

https://orcid.org/0000-0002-8599-3712

deep-learning-based play identification methods have

been proposed, in which use of CNNs (convolutional

neural networks) is the basic approach to achieving a

certain level of precision, where the output from the

CNNs is fed to other computer systems that are

capable of performing the play identification tasks

even at a more advanced level. Therefore, in this

study, motion identification experiments using a

CNN was conducted in order to achieve a certain

level of identification precisions that is needed of the

CNN that gets connected to a system that performs

advanced identification of various motions. The study

also compared its findings to the result of some

previously conducted motion identification

experiments that tested the subjects wearing sensors

on their bodies, and also examined the efficacy of the

methods it suggests.

Sato, K., Matsubara, H. and Suzuki, K.

Commonality of Motions Having Effective Features with Respect to Methods for Identifying the Moves Made during Kumite Sparring in Karate.

DOI: 10.5220/0010132400750082

In Proceedings of the 8th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2020), pages 75-82

ISBN: 978-989-758-481-7

2 RELATED STUDIES

When categorizing published studies conducted on

methods for identifying the motions of humans in

various sports and martial arts, they can be roughly

divided into two groups, i.e, studies on methods that

involve the subjects wearing sensors on their

bodies(Kwon and Gross,2005;Kwon et al.,2008), and

studies where the subjects don’t wear any sensors. As

it’s impossible for the subjects to wear any sensors in

the case of kumite competition in karate, this study

mainly deals with published studies on methods that

don’t involve the subjects wearing any sensors. Such

methods can be further broken down into

subcategories based on the types of sensor technology

they use, which are the method that uses RGB-D

sensors, the method that employs LIDAR (light

detection and ranging), and the method that utilizes

RGB cameras.

Concerning the method that uses RGB-D sensors,

one notable published study that was conducted using

the Microsoft Kinect sensors to identify the motions

of karate (Hachaj et al., 2015). However, this

particular study only measured the basic kicking and

defending motions of one subject from the front, and

so should be deemed basically the same as a number

of other research papers that only dealt with simple

motions.

As for the studies that focused on the method

employing LIDAR, one reports its application to a

scoring system used in gymnastics (Sasaki,

2018;Tomimori,2020). More specifically, the study

improved LIDAR technology so it could be applied

to 3D laser sensors for use in gymnastics competition,

which enabled precise measurement of human

motions. This system developed in the study was able

to use machine learning to recognize a skeletal model

off of depth images and automatically scored each

performance put on by a gymnast according to

scoring criteria that utilized the skeletal model.

In regard to the studies conducted on the method

using RGB cameras, in one published study, captured

video data was divided up into still images, and the

pose estimation library OpenPose was used to

generate a human skeletal model from those still

images. Then, data on the features of the skeletal

model wad fed to a neural network for learning, and

then motion identification was performed (Nakai et

al.,2018;Takasaki et al., 2019). In addition, there are

other published studies that used deep learning to

recognize various types of plays made in athletic

competition, including a study done on tennis (Mora

and Knottenbelt, 2017), a study done on ice hockey

(Tra et al., 2017), a study done on volleyball (Ibrahim

et al.,2016) and a study done on football (Tsunoda et

al., 2017). When using a method that utilizes deep

learning to identify different plays that occur in

athletic contests, it’s important to achieve a certain

level of precision using a CNN, where the output

from the CNN is connected to a system capable of

identifying various plays at a more advanced level.

The foregoing is a list of key published studies on

motion identification that don’t involve the subjects

wearing any sensors on them. Meanwhile, when it

comes to the research methods that involve use of

RGB-D sensors and LIDAR or OpenPose, motion

data is measured based on a human skeletal model.

However, in the case of kumite contests in karate, the

nodes (joints) of each human skeletal model being

used to measure the subjects’ motions might get lost

with high frequency, which is a major issue. To

address it, this study adopted a method that identified

the subjects’ various motions based on images

captured in video data that were acquired with RGB

cameras.

Figure 1: Base Labels Edited into Learning Labels for a

CNN.

3 EXPERIMENT SYSTEM

3.1 Labelling and Suggested Method

Labels 0 through 14 and 99 as specified are assigned

to the corresponding image frames extracted from the

Offensive impact

Start of an

attackin

motion

End of an attacking

motion

Image frames

0 0 2 2 2 2 2 12 2 2 2 2 2 0

0 0 0 0 1 1 1 1 1 1 1 0 0 0

Editing

Labels for CNN learnin

Impact width

Boundary label Boundary label

icSPORTS 2020 - 8th International Conference on Sport Sciences Research and Technology Support

captured video data. These label values are not ones

to be fed to a CNN for learning but are created for the

purpose of recording the subjects’ motions, which are

referred to as “base labels” in this study. As shown in

the illustration of a front kick in Figure 1 below, and

so on. The moment either subject’s attacking limb

lands on the opponent is referred to as an “offensive

impact” in this study. After the base labels are created

as described above, they are edited to suit the purpose

of each experiment, and then the label values used for

CNN learning are created.

3.2 Overview of the Experiment

System

Figure 2: Outline of the Experiment System.

An overview of the experiment system t used in this

study is shown in Figure 2. Various karate motions

are captured using a low-cost video camera for home

use, after which the video data is divided into still

images at a rate of 30 frames / sec. Then, the still

images are assigned the base labels that match them

as specified, using the application software developed

by our laboratory for motion identification

experiments.

Then, those base labels are converted into matching

label values through editing for CNN learning in a

format suitable for the purpose of each motion

identification experiment as shown in Figure 1, and

then the label values and the still images are fed to a

CNN for learning. In terms of how the CNN was set

up in this study, Keras, which has TensorFlow on the

back end, was used. Keras is a library that allows

creation of neural networks with relative ease. As for

the design of the CNN used in this experiment, it has

a structure typically seen in image recognition, with

the first part of the neural network being comprised

of convolutional layers and pooling layers, with the

latter part consisting of fully connected layers and

dropout layers. As for the algorithm for updating the

neural network weights, Adam was used in this study,

while ReLU was used as the activation function.

4 RESULT

4.1 Punch-only Yakusoku Kumite

(Pre-arranged Sparring)

To test the efficacy of the method suggested in this

study, a motion identification experiment was

conducted, in which the subjects engaged in yakusoku

kumite (pre-arranged sparring) that only allowed

punching. This experiment had two subjects, one of

which had no experience with karate, and the other a

skilled karate practitioner. In this experiment, a total

of two video cameras were used to capture the

subjects’ images from the front, from stationary

positions at a 45-degree angle from both sides. The

two subjects then took turns throwing punches at each

other, and traded their left-right positions from one

round to the next, as viewed from the cameras. When

any of the punches either subject threw so much as

touched the opponent (i.e., “skin touch”), it was

deemed effective. Meanwhile, if a punch thrown by

either subject was met with a defensive technique of

the opponent or if the opponent avoided the punch by

distancing, it’s deemed ineffective. In the experiment,

a total of 1,600 attacking motions occurred, as broken

down in formula (1) specified below.

20 reps × 5 sets

× 2 (left and right hands)

× 2 (effective/ineffective motions)

× 2 (positions) × 2 (subjects)

= 1,600 attacking motions (1)

The test data considered in the experiment was

sampled by taking one set’s worth of data from each

round that took place under different conditions as

shown in formula (2) below, which resulted in a

sample size of 20% of all data.

20 reps, 1 set × 16 rounds

= 320 test motions (2)

The experiments as indicated in Tables 1 through

3 were all conducted with an epoch size of 1,500. In

terms of what the label values specified in Tables 1

Application software for motion

identification experiments

・Image viewer

・ Label creation and

editing

・ Verification of AI-

generated

extrapolation

Video data

Keras・TensorFlow

Still images

Excel sheet, labels

Commonality of Motions Having Effective Features with Respect to Methods for Identifying the Moves Made during Kumite Sparring in

Karate

and 2 mean, numbers 1.1 through 1.6, 2.1, and 2.2

having a label value of 0 means it’s not an attacking

motion, and if any of their label values is 1, it means

it’s an attacking motion. So in these experiments,

each motion is determined either as an attacking one

or not, based on inference. If numbers 3.1 and 3.2

have a label value of 1, it’s deemed an ‘effective

strike (hit the opponent)’ while them having a label

value of 2 means they are ‘ineffective strikes (not

hitting the opponent).’ In these experiments, the

subjects’ attacking motions were identified in further

detail.

Table 1: Accuracy Rate in Relation to Impact Widths.

No. of

frames

comprising

impact

width

Rate of accuracy of label

value per image (%)

0 1

1.1 1 99.3 34.2

1.2 3 98.7 60.3

1.3 5 98.2 73.7

1.4 7 96.9 76.8

1.5 9 96.4 72.9

1.6 11 95.4 71.9

Table 2: Accuracy Rate in Relation to Boundary Label

Values.

Boundary

label

value

Rate of accuracy of label

value per image

(%)

Rate of

accuracy

attacking

motions

(%)

0 1 2

2.1 99 98.3 89.6 93.1

2.2 0 98.0 71.3 87.8

3.1 99 98.6 90.3 85.1 93.1

3.2 0 97.6 81.8 63.6 84.3

Table 3: Distance Until Offensive Impact.

Distance until offensive

impact

Method 1 Method 2

2.1 0.2 1.7

2.2 0.5 1.6

3.1 0.2 1.6

3.2 0.4 1.6

While a label value representing each attacking

motion was assigned based on its offensive impact in

a manner that maintained front-back symmetry, such

width of frames as shown in Figure 1 is referred to as

“impact width” in this study.

When the value of impact width was 7 in the

experiment described in Table 1, the rate of accuracy

on the motions assigned a label value of 1 was the

highest. Therefore, the following experiments were

all conducted with an impact width of 7.

As shown in Figure 1, the labels that are assigned

to the image frames extending from the edges of the

impact width to the start and end of each motion are

referred to as “boundary labels” in this study. With

this in mind, an experiment was conducted on

motions having boudary label values of 0 (non-

attacking motion) and 99 (outside the learning scope)

as specified in Table 2. When motion data with a

boundary lable of 99 were compared to those with a

boundary label value of 0 were compared, the

accuracy rate on label value per image improved by

18% with the label value being 1 as specified in 2.1,

and by 22% with the label value being 2 as specified

in 3.2.

The accuracy rate was close to 100% on the

motion data having a label value of 0 (non-attacking

motion) in each experiment. This could be attributed

to the fact that all attacking motions could be detected

almost entirely within their impact width. When the

extrapolation process occurring on the video data was

checked using the viewer, the extrapolation result on

each attacking motion made could be confirmed near

the offensive impact, which is quantitatively

expressed in the columns “Method 1” and “Method 2”

under “Distance until offensive impact” in Table 3.

Based on these data, it’s possible to know when an

extrapolation is made on each attacking motion in

terms of how close it is to the offensive impact.

Method 1

Distance＝

｜‘Avg. frame numbers accurately

extrapolated by the CNN’

－ ‘frame numbers corresponding to the

offensive impact’｜ (3)

Method 2

Distance=Avg. of {｜‘frame numbers

accurately extrapolated by the CNN’

－ ‘frame numbers corresponding to the

offensive impact’｜} (4)

Method 1 is used as a means by which to

extrapolate the offensive impact using the CNN. It

calculates the difference between the arithmetic mean

value of the frame numbers that the CNN accurately

icSPORTS 2020 - 8th International Conference on Sport Sciences Research and Technology Support

extrapolated within the impact width, and the frame

numbers corresponding to the offensive impact. As

the result of the extrapolation performed by Method

1 does not exceed 0.5 as shown in Table 3, it’s

discernible that it almost perfectly captured the

frames of each offensive impact observed. However,

there are some instances where the CNN-inferred

frame numbers deviated from the intended offensive

impact by large margins, on either side, while there

are other instances where the calculated mean values

happen to fall on the frame numbers that are in the

centers of the offensive impacts observed. Meanwhile,

in the case of Method 2, the mean values are

calculated after calculating the distance between the

subjects by each inferred frame number, the distance

until offensive impact is accurately represented. From

the result of the extrapolation made using Method 2

as specified in Table 3, it’s discernible that Method 1

＜ Method 2.Based on this observed relation between

the two methods, it’s apparent that the CNN-inferred

non-zero label values representing the impacts are

distributed being centered on the offensive impacts.

Concerning the “rate of accuracy on attacking

motions (%)” as specified in Table 2, each instance

where the CNN’s inference was accurate within the

impact width is deemed as an accurate inference for

each attacking motion concerned. This adjustment

was made because there were many instances where

the rate of accuracy of label value per image was low

but the offensive impact was recognized correctly

even if only one frame was inferred accurately.

This is also supported by the fact that the values

indicating the distance until offensive impact have

been small. The rates of accuracy on attacking

motions for 2.1 through 3.2 specified in Table 2 were

highly precise, ranging between 84.3％ and 93.1％,

indicating the efficacy of the method suggested in this

study.

4.2 Yakusoku Kumite (Pre-arranged

Sparring) Simulating Actual

Competition

An experiment was conducted to identify the subjects’

motions in yakusoku kumite (pre-arranged sparring)

simulating actual competitive matches, in which both

punches and kicks were allowed. The subjects’

images were captured using four units of overhead

cameras, at a height of 5.75 m above the floor. In this

report, the motion identification experiment was

conducted based on two cameras’ worth of data, so

that it could be compared to another experiment done

on punch-only kumite matches, which was recorded

using two cameras also. Recording the video data

from overhead positions eliminated the possibility of

any blind spots occurring due to the main referee

blocking the view, preventing all motions of the

contestants to be seen clearly in official karate

matches. Such camera positioning would also reduce

the likelihood of contestants blocking the view to

ensure accurate judging.

For this experiment, the parts of the subjects’

bodies that would be involved in the moves they

would make on each other and the order of the moves

were decided in advance. The subjects were also told

beforehand that they could through punches in one-

two combinations, and also kick both middle and high.

The subjects were allowed to move freely inside the

designated court, and could initiate their attacking

motions at any time they would like. As per the

Olympic rules, each effective attacking motion had to

be either a light touch or a non-contact strike stopping

several centimeters short of contact. Meanwhile, an

attacking motion was deemed ineffective when it was

successfully defended by the opponent by deflection,

evasion, or distancing. As for the composition of the

subjects, a total of six high school students all having

a 1

-degree black belt took part in kumite matches in

three pairs, the first pair being subjects A (red) and B

(blue), the second being subjects C (red) and D (blue),

and the third being subjects E (red) and F (blue). The

motions of the subjects that were supposed to occur

within one round of kumite are described in (a) and

(b) below.

(a) Left and right punches had to be thrown in that

order. They could be thrown in series or in

sporadic bursts. Such one-two punch

combination had to be thrown for a total of 20

times, and subject ‘red’ had to go first, followed

by subject ‘blue’ each time.

(b) Kicks also had to be thrown by the subjects

following the same rules applied to punches.

The total number of attacking motions executed

by the subjects for both (a) and (b)＝160 / round.

While following the round-by-round rules

described above, the subjects sparred for the numbers

of rounds as specified below.

Effective attacking motions (non-contact): 10 rounds

Non-effective attacking motions (misses): 10 rounds

Total number of attacking motions attempted

＝160 / round × 20 rounds

＝3,200 (5)

(no. of punches: 1,600; no. of kicks: 1,600)

Commonality of Motions Having Effective Features with Respect to Methods for Identifying the Moves Made during Kumite Sparring in

Karate

Table 4: Rate of Accuracy on Attacking Motions.

No Subject

Rate of accuracy of label

value per image (%)

Rate of

accuracy

attacking

motions

(%)

0 1 2

4.1 AB 97.9 66.8 94.2

4.2 CD 98.5 46.5 85.6

4.3 EF 95.0 70.6 92.6

5.1 AB 97.9 63.9 62.2 91.5

5.2 CD 99.1 35.2 40.7 77.7

5.3 EF 94.0 64.5 65.0 89.1

Each pair was assigned a total of six rounds,

consisting of three rounds of effective attacking

motions and three rounds of non-effective attacking

motions, plus some extra work α depending on the

pair.

Extra workα: The pair A and B was assigned one

additional round of effective attacking motions while

the pair C and D was assigned one additional round

of non-effective attacking motions.

Table 5: Distance until Offensive Impact.

No Distance Until Offensive Impact

Method 1 Method 2

4.1 0.7 1.9

4.2 1.0 1.9

4.3 0.7 1.9

5.1 0.8 1.9

5.2 1.1 1.8

5.3 0.8 1.9

Test data on a total of two rounds contested by

each pair, consisting of one round of effective

attacking motions and one round of non-effective

attacking motions, were used. While a total of 20

rounds took place, the number of rounds that made up

10% of all rounds were used to extract test data from.

The result of the test executed is shown in Table 4.

The number of epochs was set at 500.

The rate of accuracy of label value per image on

all motions of which the label value was 0 (non-

attacking motions) for numbers 4.1 through 5.3 was

fairly high, ranging between 94.0% and 99.1%. As for

the distance until offensive impact as specified in

Table 5, it turned out be less than two frames. This

means that had either a label value of 1 or 2 centering

on the offensive impact frames occurring within the

impact width were detected, which was also

confirmed during an observation that used the viewer.

Figure 3: No 4.1, Losses During the Learning Process, and

Motion Identification Accuracy.

Meanwhile, the rate of accuracy of label value per

image for the label values 1 and 2 turned out to be

slightly low. However, it must be noted that this rate

of accuracy does not indicate the rate of accuracy on

the attacking motions attempted. As far as the method

suggested in this study is concerned, what’s important

is to accurately capture each offensive impact that

occurs. Therefore, even in instances where only one

image frame within any given impact width is

accurately inferred, the motion identification is

deemed to have functioned correctly on those

attacking motions per se. Hence, in this experiment,

if the CNN extrapolation is done correctly within the

impact width of a motion, the extrapolation is deemed

accurate on the motion, since its offensive impact is

captured almost completely, as was the case with the

punch-only kumite. The average rate of accuracy on

attacking motions for numbers 4.1 through 5.3 turned

out to be quite high at 88.5％.

Figure 3 contains graphs indicating the losses that

occurred during the learning process of number 4.1

specified in Table 4, along with the corresponding

motion identification accuracy. 10% of the data was

unknown data not part of the training data, which was

intended for testing use.The losses apparently

icSPORTS 2020 - 8th International Conference on Sport Sciences Research and Technology Support

increased as the learning process progressed, while

there is a sign of overtraining. Such patterns could

also be discerned from the experiment rounds

numbered 4.2 through 5.3.

5 DISCUSSION

In the experiment covering numbers 1.1 through 1.6

specified in Table 1, the rate of accuracy is the highest

on the motions having a label value of 1 (attacking

motions) when the impact width is 7. In addition, as

evident in the results shown in Table 2, the rate of

accuracy of label value per image improved by

roughly 20% when the boundary label was 99. This

might mean that there were motions with features that

would prove effective for motion identification that

exit within a certain range centering on the offensive

impact, while there were those other motions

immediately preceding and following the

aforementioned range having features that lowers the

motion identification accuracy. Phenomena similar to

these were also encountered in a separate experiment

conducted previously that used optical motion

capture (Sato and Kuriyama, 2011). It’s believed that

the motions having effective and non-effective

features might be the same as the motion

identification that uses the joint position data

generated with a human skeletal model, and also as

the motion identification based on image data.

As the fairly high rate of accuracy of about 90%

could be achieved on attacking motions as specified

in Table 4, and given how accurately the offensive

impacts that occurred in each experiment could be

captured, it could be surmised from a comprehensive

viewpoint that the desired level of CNN-assisted

motion identification accuracy, which is the objective

of this study, might have been achieved, which is

necessary for connecting to a system capable of

identifying various motions at a more advanced level.

6 CONCLUSION

As there has been no published study on basic

methods for acquiring motion data and for identifying

various motions in karate kumite competition, this

study conducted CNN-assisted motion identification

experiments that acquired data using overhead video

cameras placed above the contestants so that they

wouldn’t interfere with the contests in progress.

CNNs such as one used in this study are connected to

advanced sports-specific extrapolation systems such

LSTM, on which there have been published studies

focusing on other sports

(Tsunoda et al., 2017). While

each CNN to which the aforementioned connection is

made must possess a certain level of motion

identification accuracy, sufficient results have been

achieved during the experiments conducted in this

study.

Explained above is the basic method that this

study suggests for effective data acquisition and

motion identification applicable to karate kumite

contests. In terms of what could be improved in the

future, it will be necessary to check the efficacy of the

method when the number of cameras is increased, and

also to conduct experiments on the identification of

motions that more closely resemble the contestants’

actual movements in official competitive matches.

REFERENCES

Hachaj, T., Marek R. O., and Katarzyna K. (2015).

Application of Assistive Computer Vision Methods to

Oyama Karate Techniques Recognition. Symmetry,

1670-1698.

Ibrahim, M. S., Muralidharan, S., Deng, Z., Vahdat, A., and

Mori, G. (2016). A Hierarchical Deep Temporal Model

for Group Activity Recognition. Proceedings of the

IEEE Conference on Computer Vision and Pattern

Recognition (CVPR), 2016, pp. 1971-1980.

Kwon, D. Y., and Gross, M. (2005). Combining Body

Sensors and Visual Sensors for Motion Training.

The 2005 ACM SIGCHI International Conference on

Advances in computer entertainment technology.

Kwon, T., Cho, Y., Park, S. Il., and hin, S. Y. (2008). Two-

Character Motion Analysis and Synthesis. IEEE

Transactions on Visualization and Computer Graphics,

14(3), 707-720.

Nakai, M., Tsunoda, Y., Hayashi, H., and Murakoshi, H.

(2018). Prediction of Basketball Free Throw Shooting

by OpenPose. JSAI International Symposium on

Artificial Intelligence, 435-446.

Mora, S. V., and Knottenbelt, W. J. (2017). Deep Learning

for Domain-Specific Action Recognition in Tennis.

2017 IEEE Conference on CVPRW, 170-178(online).

Sasaki, K. (2018). 3D Sensing Technology for Real-Time

Quantification of Athletes' Movements. Fujitsu, 13-20

in Japan.

Sato, K., and Kuriyama, S. (2011). Classification of karate

motion using feature learning. 2011 by Information

Processing Society of Japan, 75-80 in Japan.

Takasaki, C., Takefusa, A., Nakada, H., and Oguchi, M.

(2019). A Study on Action Recognition Method with

Estimated Pose by using RNN. 2019 Information

Processing Society of Japan in Japan.

Tomimori, H., Murakami, R., Sato, T., and Sasaki, K.

(2020). A Judging Support System for Gymastics Using

3D Sensing. Journal of the Robotics Society of Japan in

Japan.

Commonality of Motions Having Effective Features with Respect to Methods for Identifying the Moves Made during Kumite Sparring in

Karate

Tra, M. R., Chen, J., and Little, J. J. (2017). Classification

of Puck Possession Events in Ice Hockey. 2017 IEEE

Conference on CVPRW, 147-154 (online).

Tsunoda, T., Komori, Y., Matsugu, M., and Harada, T.

(2017). Football Action Recognition using Hierarchical

LSTM. IEEE Conference on CVPRW, 155-163.

icSPORTS 2020 - 8th International Conference on Sport Sciences Research and Technology Support