The Development of Automatic Training Analysis Using 3D

Accelerometer in Male Young Elite Soccer Team

Takuya Magome

1,2 a

,Takeshi Tanaka

and Toru Yamaguchi

Otemon Gakuin University, Osaka, Japan

Department of Health and Sport Sciences, Osaka University Graduate School of Medicine, Osaka, Japan

Hitachi, Ltd. Research & Development Group, Tokyo, Japan

Keywords: 3D Accelerometer, Automatic Training Analysis, Injury Prevention, Sports Support.

Abstract: In recent years, there has been a growing need to introduce IT technology to improve performance in

professional sports and athlete training. However, while technological innovation is taking place, the amount

of data that can actually be acquired is becoming enormous, and unless there are people who can provide

expert advice, the compatibility with training is not high. In this study, we aimed to build a system that uses

a three-dimensional accelerometer to appropriately evaluate the performance of athletes during training. We

developed a classification function that detects and separates intervals from the measured time series data and

automatically divides the separated sections by type of training. As a result of evaluation using data from 16

days of training for a U-18 soccer team (n=33), intervals were detected with 100% accuracy, and five types

of training were detected with 80% accuracy. We confirmed that a system equipped with the developed

function could speed up feedback to athletes.

1 INTRODUCTION

1.1 Background

In the field of professional sports and athlete training,

there is a growing need for the introduction of IT

technology for the purpose of performance

improvement and injury prevention. In the past,

evaluation of individual athletes' performance and

team tactics and strategies in each sport was done

subjectively by visual observation and video analysis

by the coaches and instructors involved in the

respective sports. Currently, with the development of

highly accurate and compact wearable sensors and

video analysis (Vickery et al.,2014). it is becoming

possible to collect objective and quantitative data at

all times during daily training.

Dhruv et al. (Dhruv et al.,2019) have shown that

comparing the intensity of exercise during the week

in question with the intensity of exercise in the past

can lead to an estimation of the risk of injury to the

athlete. The objective evaluation of the conditioning

https://orcid.org/0009-0008-3105-8869

https://orcid.org/0000-0002-1178-837X

of athletes without relying on subjectivity is very

useful not only for the individual but also for the team.

In addition to measurement technology, it is also

expected to prevent injuries caused by overtraining by

constantly monitoring the condition of players in

various sports during games and practices and

providing real-time feedback to the staff and players

on site. Among IT technologies, wearable sensors are

particularly suited to grasp actual numerical

behaviours in various competitions because they can

measure physical activity and vital conditions by

simply being worn.

However, the burden of analysis is heavy on the

field staff and others, and time is required for analysis

after games and practices. For example, time is

required to organize and manage the obtained data by

linking them to the practice contents, and to adjust the

analysis methods to the team characteristics. Although

it is important to analyse the relationship between the

level of physical activity and the occurrence of

traumatic injuries and disabilities in sports and to use

this information for prevention, it is still difficult to

immediately predict the risk in the field.

160

Magome, T., Tanaka, T. and Yamaguchi, T.

The Development of Automatic Training Analysis Using 3D Accelerometer in Male Young Elite Soccer Team.

DOI: 10.5220/0012940700003828

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 12th Inter national Conference on Sport Sciences Research and Technology Support (icSPORTS 2024), pages 160-167

ISBN: 978-989-758-719-1; ISSN: 2184-3201

In addition, in the case of exercise data, for

example, a decrease in running distance can be

considered as a decrease in the condition of the athlete.

However, it is not easy to distinguish which factor is

responsible for the decrease in running distance,

because the values change depending on the

conditions and types of training menus that athletes

are required to follow, regardless of their own

conditions.

To achieve this, it is necessary to view the obtained

data instantly and in an easy-to-understand manner. It

is also necessary to provide numerical values that can

be easily conveyed to athletes even by staff who do

not have much specialized knowledge. Therefore, this

research focused on developing equipment and

measurement systems that meet the needs of the field.

Therefore, in this research, the time required for

data analysis is greatly reduced by automating the

organization of acquired data by dividing the time

according to the content of practice by introducing AI

analysis technology for sensor and video data and

past injury history, etc., which has been difficult in

the past, and by detecting signs of injury in real time

and adjusting the amount of practice, which have

been difficult in the past. The system aims to detect

signs of injury in real time and to provide feedback

such as adjusting the amount of practice, which was

difficult in the past.

1.2 Purpose of Research

This study aims to reduce the burden of on-site data

analysis, we aimed to develop technology that

automatically classifies practice types from data

collected using wearable acceleration sensors and to

verify its accuracy.

2 MATERIALS AND METHODS

2.1 Performance Analysis Using IT

Technology

This study clarifies the elements necessary for young

players to make their professional debut and play

soccer as a professional contract player not only in

Japan but also in other countries. By analyzing in

detail what kind of training is being done on a daily

basis, we hope to give the players an opportunity to

think about it. We will clearly and visually appeal to

them what kind of training will produce what level of

exercise intensity. In addition, teams with abundant

financial resources can purchase and wear famous IT

equipment, but if they are not economically well-off,

there are fewer opportunities to wear cutting-edge

equipment. In response to these, using our own

developed sensors will also help eliminate economic

disparities in sports. To achieve this, data from games

and practices must be analyzed quickly. It is

important to acquire data in real time and use it to

protect players. Therefore, we devised a system that

not only measures data but also provides real-time

feedback as shown in Fig 1.

Figure 1: Packaged data feedback system.

2.2 Male Young Elite Soccer Player

The subjects were 33 youth soccer team players from

the Japan Professional Football League (J-League)

(teams with experience in the top ranks of

international youth soccer tournaments), aged 15-18

years old. The period of the study was 16 days,

excluding days when practice was suspended and

days of international matches. The average practice

time was 100 minutes/day, with a maximum of 2

hours/day and a minimum of 1.5 hours/day. Since

some of these youth players will be promoted to the

The Development of Automatic Training Analysis Using 3D Accelerometer in Male Young Elite Soccer Team

161

top team and become candidates for the national

team, the data from their youth period is very

important.

2.3 Measured Training

The content of the exercises was confirmed with the

team coaches and consisted of five types of training:

warming up and basic training (including stretching),

group training (in which the amount of exercise

varied greatly from player to player), low (Tactics

training), medium, and high (Game training) intensity

training in which all the players had the same amount

of exercise. The training consisted of five types of

training, which could be classified into low (Tactics

training), medium and high intensity (Game training).

The composition of the training sessions (Day 11 and

Day 15) and the start and end times of each session

were recorded visually and on video.

2.4 Wearable Acceleration Sensors

As shown in table1, a wristwatch accelerometer UW-

301BT manufactured by A&D was used to measure

athletes' exercise intensity. Figure 2 shows that the

sensor was worn on the opposite wrist of the

dominant hand, the measured data was triaxial

acceleration, the sampling rate was 20 Hz, the

measurement range was -4~+4 G, and the oxygen

uptake at rest of 3.5 (mL/kg/min) was set as 1 G. It

was also confirmed that the device could be used

without any load or discomfort to the athlete and

without interfering with the athlete's movement or

training.As an index to evaluate the athlete's exercise,

we scalarized the 3-axis acceleration vectors and

calculated the averaged value in seconds as the

exercise intensity (unit:G). By averaging over

seconds, changes in acceleration due to minute arm

movements can be smoothed out and converted into

values that reflect whole-body movements, thus

reducing the amount of information handled to the

minimum necessary and facilitating analysis.

Table 1: Specification of wearable device.

Figure 2: Appearance of wearable device.

2.5 Methods

In order to reduce the burden of data analysis in the

field, which is the purpose of this study, we first

aimed to detect intervals during practice by

automatically dividing the data into practice types.

Using the characteristic that the waveform of the

exercise intensity of all the athletes decreases during

intervals, we extracted the time period when the

average value of the exercise intensity of the athletes

performing the same practice for a certain period of

time was less than a predetermined threshold value.

The threshold value and the length of time for

calculating the average value were determined by

searching for the value that would increase the

accuracy of interval detection the most (see table2).

The classification of practice content was then

calculated from the athlete's exercise intensity data

for each interval separated by intervals. Using

multiple features corresponding to each interval, we

used the K-means clustering method (Andri, 2022;

Luiz et al, 2017), one of the popular unsupervised

clustering methods, to mechanically classify the data

into a pre-defined number of clusters. In general

classification algorithms such as K-means clustering

methods, it is necessary to select features that clearly

differ by practice content in order to obtain the

desired classification accuracy. Based on the

observation of the practice contents and the

investigation of the practice contents of general group

sports, we hypothesized that the magnitude of players'

movements (maximum/average), the similarity of

movements among players (small variation), and the

uniformity of exercise intensity within a segment

(small variation) would be useful features, and

designed a set of features shown in (see table3).

The proposed method was applied to the collected

data for 14 days to detect intervals and classify each

interval. For the classification, the number of

classifications (see table3), which indicates the type

of practice, is set to 4, since the team has roughly four

types of training. The detected intervals and the

icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support

162

Table 2: Interval detection (study by exercise intensity and time).

Table 3: Practice menu detection by K-means clustering method.

classification results were compared with the records

of the two days, and the accuracy was evaluated by

the rate of agreement.

3 RESULTS

In the detection of practice intervals, the interval

duration and threshold values were defined by

observation and investigation of actual training and

analysis of exercise intensity data. The specific values

for the threshold values were set by using the values of

Day 11 and Day 15, which were the teacher data within

the measurement period, and by comparing the

exercise intensity with the records of the content of the

training for the two days by combining the time of 20,

40, 60, and 80 seconds for the four types of exercise

intensity of 1.5, 2, 2.5, and 3 METs. We searched for

the values that resulted in a 100% positive predictive

value and a 0% false positive rate (see table 2). As a

result, the categories of 2.5 METs and 1 minute were

found to be the most consistent with the actual intervals.

The Development of Automatic Training Analysis Using 3D Accelerometer in Male Young Elite Soccer Team

163

In general, during intervals, athletes often listen to

instructions from their instructors, in addition to

rehydrating and preparing for the next training

session. These criteria are also considered to be

appropriate in that, based on the same observations,

the threshold values are considered to be appropriate

for exercise intensities that indicate walking or

standing still. Fig 3 shows, with reference to these

results, the exercise intensity data of Team A, which

consists of 16 members, the team with the highest

level among 33 members, for one day of practice. The

waveforms representing the exercise intensity are

also consistent, and as mentioned above, it is

suggested that the interval is also the place where the

intensity drops all at once.

Next, the classification of practice content was

calculated from the athlete's exercise intensity data

for each interval (see table 3). The results of the

detection of intervals and the classification of practice

are shown in Fig 4. Especially for the two days with

visual recordings (Day11 and Day15), we compared

the classification results with the results of visual and

video recordings. The validity of the results of the

classification of the training menus was also

examined by interviewing the coaches and by visual

recording. The results of applying the classification

of the automatically clustered training menus to the

classification of actual training sessions are also

shown, based on the comparison with the values of

the feature values and the observation recordings.

The training menus automatically clustered by the

developed process were classified into 5 levels from

Training A to E based on the values of the feature

values (see table3). In comparison with the

observation records, we found that the classification

of Training A (Warming up), Training B (Group

training), Training C (Game training), Training D

(Tactics training), Training E, and the actual training

menus are one-to-one. Training A (Warming up),

Training B (Group training), Training C (Game

training), Training D (Tactics training), Training E,

and the classification of actual training menu. On the

other hand, the classification E was found to indicate

a rest interval that was not detected as an interval or a

training session with particularly low exercise

intensity. This may be due to the fact that physical

training was not conducted during the measurement

period, resulting in a different classification from the

actual one.

Next, the results of the classification of training

menus were compared in detail with the observation

records to evaluate the accuracy of the classification.

We found that all of the training breakpoints were

consistent with those in the observation records.

Furthermore, the classification of the training menus

was found to be consistent with the observed data

with a precision of 9/11. In particular, group training,

including 4-on-2 ball keeping, which is the most

frequently practiced exercise by the subject teams and

is performed almost every day, was consistent

without error. On the other hand, warming up and

basic training, which are generally less intense, were

difficult to distinguish from the other categories,

especially when a more intense element was added.

Figure 3: Exercise Intensity per Day.

Figure 4: Interval detection and practice classification.

4 DISCUSSION

In this study, based on the data obtained from the

results of practice and games with the wristband-type

icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support

164

wearable accelerometer, a feedback report was

prepared on the exercise intensity of each player

regarding the evaluation items, and a report was made

to the team. However, since each player and coach

interpreted the measured data, there were cases in

which it was difficult to understand the numerical

values obtained and the implications derived from

them unless a researcher was present. Even if an

existing measurement system is introduced, it is not

uncommon in the field of sports such as soccer that

the data are not actually utilized in training, and the

difficulty of how to process the huge amount of data

obtained and how to make sense of them and how to

understand the numerical values and utilize them in

the field is also a problem that needs to be addressed.

The challenges are not limited to the field of sports,

but include the difficulty of how to make sense of the

huge amount of data obtained and how to understand

and utilize the numerical values. We have to give due

consideration to this issue because it is related to our

own health care (Mowafa & Bakheet,2017).

Therefore, we referred to specific advice from team

doctors, trainers, and coaches, as well as surveyed the

opinions of athletes who actually took the

measurements, and extracted usefulness and issues

related to the report content and evaluation items. In

addition, in the description of exercise intensity, the

measured acceleration values (unit:G) are not directly

described, but are converted into METs -Metabolic

equivalents, (Jette et al.,1990) which is a standard to

indicate the amount of physical activity. METs are

based on energy expenditure, and the correspondence

table between specific physical activities and METs

conversion is also available, making it easy to

intuitively understand the intensity of exercise. To

convert acceleration into METs, three subjects (age

39.7±1.5 years; height 180.0±7.9 cm; weight

79.6±12.6 kg) were asked to wear a prototype with

the same performance as the UW-301BT and

Suzuken's Lifecoder EX at all times except when

taking a bath for one month. The absolute values of

the 3-axis acceleration measured by the UW-301BT

were summed up in correspondence with the 9 levels

of METs (Kumahara et al.,2004) that can be

calculated from the LifecoderEX data, and the

average value per unit time (minutes) was calculated.

Regression analysis using these nine data points

yielded a coefficient of determination of R

=0.928

(p<0.001). As a result, the results can be confirmed at

a glance with easy-to-understand indicators, as in the

feedback report reported in the previous issue. Since

the analysis with utility is also possible, G notation is

also available upon request, as shown in Fig 4, in

response to the needs from the field. In order to

realize the quantitative condition evaluation by the

wearable sensor, the policy of the team conducting

the actual measurement and the numerical values and

items to be examined are likely to change, so it is

necessary to continue to give sufficient consideration

to how to feed back the results to the field. As shown

in fig.3, the waveforms of the exercise intensity of all

the athletes show that the exercise intensity increases

during the training and decreases during the interval

when the athletes move to the next training. In other

words, it is necessary to detect the interval time when

the exercise intensity decreases in order to delimit the

time of the training menu itself. However, even when

the same training menu is performed by a team, it is

possible that there are variations in the exercise

intensity of each individual, and this becomes more

pronounced in the case of injury

(Gabbett,2016,2018).

Depending on the menu, it is possible that not all

the participants necessarily participate in the training,

but only some of them do, or that they take a break.

Therefore, we averaged the waveforms of the

exercise intensity of all the athletes who were

performing the same pattern of training menu, and

then we set up a break in the training menu. This

approach is considered to have reduced the influence

of individual variations and enabled more accurate

automatic detection of training intervals. There are a

wide variety of patterns in the automatic detection of

intervals, depending on each athletic organization.

There are differences between individual and team

competitions, and between ball games and other

sports. Therefore, we defined the characteristics of

intervals in this study by analyzing visual

observations, surveys, and video analysis of actual

training sessions. As shown in table 2, we found the

optimal interval value of 2.5 METs or less to be

detected from the exercise intensity data, and we also

examined the time during the interval as a condition.

Specifically, when the interval is shorter than 1

minute, it is often not accompanied by a change of the

training menu, such as the switching of roles of the

athletes depending on the menu, including hydration,

assuming the movement of the athletes themselves.

Considering a complete changeover, the interval

should be at least 1 minute at the shortest. Therefore,

we considered not only the intensity of the exercise

but also the duration of the exercise. Therefore, it is

necessary to consider not only the exercise intensity

but also the time in future efforts to detect intervals,

as this will increase the accuracy. In addition, the

waveform data of the exercise intensity of each

athlete obtained from each training menu was used to

classify the exercise intensity patterns of each training

The Development of Automatic Training Analysis Using 3D Accelerometer in Male Young Elite Soccer Team

165

segment by the feature values, and the K-means

clustering method was used to classify the data. In

this study, the classification was performed

comprehensively based on variables such as the

waveform of exercise intensity, length of time, and

order of exercise intensity. However, it is necessary

to develop an original algorithm depending on the

competitions actually performed within a narrow

range. Regarding the classification of training menus

(see table 3), the overall accuracy was 80%. However,

there was no error in the group training. On the other

hand, it was found to be difficult to distinguish

between warm-up and basic training, where the

intensity of exercise was generally low, and other

training sessions where a transient element of high

intensity was added, which is a subject for future

study. In this study, there was a strong need for

quantitative measurement to detect signs of injury and

prevent injury, since it is not possible for athletes to

detect signs of injury by their own judgment or that

of their coaches. The findings of this study suggest

the possibility of not only comparing the daily

exercise intensity and other conditioning data of the

same athlete, but also comparing the physical activity

level of other athletes, not on a weekly basis, but by

extracting only specific practices. This finding is not

only highly useful for conditioning the team and each

player, but also leads to the possibility of utilizing the

physical activity assessment for prevention of trauma

and injury. It is also possible to analyze effective

values for passing and ball possession practice in a

limited and confined space, and to analyze the

performance in detail. This new approach is useful

not only for training and games, but also for recovery.

It is suggested that the understanding of exercise

intensity for each training session and the

management of conditioning of each athlete, as well

as the implementation of practice, are highly likely to

lead to injury prevention.

5 CONCLUSIONS

In this study, we calculated the exercise intensity of

each individual in one second from the acceleration

data of all athletes during training. In addition, the

detection process was examined by finding the

characteristics of the breaks in the training menu. As

a result, it was possible to extract and compare the

measurement results that matched the target training

conditions, and to calculate indices such as exercise

intensity and running distance by analyzing the

measured acceleration data. Our original analysis

made it possible to automatically detect intervals in

group training, and furthermore, to automatically

classify them by training intensity. The results

suggest the possibility of utilizing the physical

activity evaluation for the prevention of trauma and

injury and for conditioning by comparing the exercise

intensity of athletes by training intensity on a daily

basis.

ACKNOWLEDGEMENTS

This research was supported by JSPS Grants-in-Aid

for Scientific Research JP21K09277 and SRIP

(Sports Research Innovation Project in Japan). The

authors thank Hitachi.co.and U-18 soccer team for

advice on this project.

REFERENCES

Vickery, William., Dascombe, Ben., Baker, John., Higham,

Dean G., Spratford, Wayne., & Duffield, Rob. (2014).

Accuracy and reliability of GPS devices for

measurement of sports-specific movement patterns

related to cricket, tennis and field-based team sports.

The Journal of Strength & Conditioning

Research.18,1697-1705.

Dhruv R, Seshadri., Ryan T, Li., James E, Voos., James R,

Rowbottom., Celeste M, Alfes., Christian A,Zorman.,

& Colin K. Drummond.(2019). Wearable sensors for

monitoring the internal and external workload of the

athlete. npj Digital Medicine.2:71

Andri Mizal. (2022). Statistical Analysis of Microarray

Data Clustering using NMF, Spectral Clustering, K-

means, and GMM. IEEE/ACM Trans Comput Biol

Bioinform. 19(2),1173-1192

Luiz,H,,Palucci Vieira., Vitor, L., de Andrade., Rodrigo, L.,

Aquino., Renato, Moraes., Fabio, A., Barbieri., Sérgio,

A., Cunha., Bruno, L., Bedo.,&Paulo, R., Santiago.

(2017). Construct validity of tests that measure

performance for young soccer players based on cluster

analysis: exploring the relationship between coaches

rating and actual measures. J Sports Med Phys Fitness.

57(12),1613-1622.

Mowafa,H., ＆ Bakheet,A. (2017). The Hazards of Data

Mining in Healthcare. Stud Health Technol Inform.

238,80-83.

Jette, M., K. Sidney., and G. Blümchen.(1990). Metabolic

equivalents (METS) in exercise testing, exercise

prescription, and evaluation of functional capacity. Clin

Cardiol. 13(8):555-65.

Kumahara,H.,Yves,Schutz.,Ayabe,M.,Yoshioka,M.,Yoshit

ake, Y., Shindo,M., Ishii, K., & Tanaka,H.(2004). The

use of uniaxial accelerometry for the assessment of

physical-activity-related energy expenditure: a

validation study against whole-body indirect

calorimetry. Br J Nutr. 91(2):235-43.

icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support

166

Tim,J.,Gabbett.(2016). The training-injury prevention

paradox: should athletes be training smarter and harder?

Br J Sports Med. 50(5):273-80.

Tim,J.,Gabbett.(2018). Infographic: The training-injury

prevention paradox: should athletes be training smarter

and harder? Br J Sports Med. 52(3):203.

The Development of Automatic Training Analysis Using 3D Accelerometer in Male Young Elite Soccer Team

167