Motion Causal Network Analysis for Quantitative Evaluation of

Baseball Form by Video Analysis

Takeshi Tanaka

and Norio Gouda

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

Hitachi, Ltd. Government & Public Corporation Information Systems Division, Tokyo, Japan

Keywords: Motion Analysis, Body Tracking, Transfer Entropy, Network Analysis, Baseball.

Abstract: In professional sports, IT technology has been introduced to strengthen players, and in the baseball industry,

where there is a large number of athletes, data analysis is expected to improve training efficiency, even for

amateurs. However, the use of IT has not been widespread in the past due to the need for special equipment

for measurement and the difficulty of interpreting data. In this study, we proposed a technique for quantifying

the interlocking nature of players' forms using the transfer entropy from time-series data of players' skeletal

coordinates obtained by image recognition to intuitively visualize the characteristics of players' forms using

only video. As a result of evaluating players' hitting form using the proposed technique, we confirmed that

the transfer entropy significantly changed in the target region when players were conscious of improvement,

and we obtained a prospect for the practical application of form analysis using video.

1 INTRODUCTION

Wearable devices, camera-based measurement

technology, and AI-based analysis technology are

rapidly being applied to sports. Major professional

sports teams such as soccer and baseball use data for

player evaluation, performance improvement, and

injury prevention. Data measurement and analysis

technologies developed in professional sports are

expected to spread to amateur sports teams and

players.

In baseball, which has long been a popular sport

in Japan, detailed scores and notes have been

manually recorded throughout professional and

amateur sports for use in competition and coaching,

and data measurement devices have been used in

amateur coaching from early on. Developed from

speed guns that use ultrasonic waves to measure the

velocity of pitches and batted balls, devices that

record not only velocity but also trajectory and

rotation in detail have become popular in professional

stadiums, and products for amateur teams are also on

the market. In addition, devices with acceleration and

gyro sensors mounted on bats and balls are also

available, enabling simple visualization and analysis

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

https://orcid.org/0000-0003-4059-220X

of pitching velocities and swing trajectories in

individual practice. In recent years, it has become

possible to take pictures of pitching and hitting forms

with a smartphone without using special equipment,

and the AI automatically recognizes joint angles and

skeletal positions from the images, converting them

into data for visualization (Chung, 2022).

However, the problem is that only a limited

number of teams and players are able to use this

technology because professional-grade measuring

equipment is expensive and requires human resources

to operate, and instructors with specialized

knowledge are needed to interpret the data and

provide guidance. Because of the high cost of

ultrasonic equipment, only top-level amateur players

belonging to famous teams can use it, and the

disparity in coaching has also been a factor hindering

the development of amateur baseball. In addition,

although measurement by smartphone is relatively

inexpensive, it has been limited to measurement of

the skeletal position, so interpretation of data has been

limited to a few instructors, making it difficult to

spread the use of smartphones among a wide range of

levels and generations of amateur baseball players.

Tanaka, T. and Gouda, N.

Motion Causal Network Analysis for Quantitative Evaluation of Baseball Form by Video Analysis.

DOI: 10.5220/0012949500003828

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

In Proceedings of the 12th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2024), pages 185-192

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

185

In this study, we aim to enable amateur athletes to

easily measure their own condition, convert it into

data, and practice PDCA in training based on the

accumulated data of analysis subjects and others by

visualizing the characteristics of pitching and hitting

form movements from videos taken with a

smartphone in an easy-to-understand manner. We

propose a "motion causal network analysis

technology" that visualizes the characteristics of

pitching and hitting form movements from videos

taken with smartphones in an easy-to-understand

manner. A method has been developed to estimate the

linkage of group movements in soccer and other

sports using transfer entropy and to visualize the

relationships among groups in a network (Itoda,

2015). In this study, we apply the method to analyze

a group and visualize the linkage by considering each

part of an individual's form as a group. In this paper,

we report on a prototype system that visualizes the

results of applying causal network analysis of

movements by inputting videos taken with a

smartphone and evaluate whether the system can

visualize the effects of the practice intended by the

individual using data from actual amateur athletes.

2 METHOD

2.1 Baseball Training Using Digital

Technology

This study aims to provide amateur baseball players

with a wide range of environments for data utilization

so that they can think and practice how to grow by

themselves. In baseball, it is common practice for

players to record their findings in a notebook and

repeat trial and error in the process of training and

receiving instruction. In recent years, with the spread

of smartphones and the Internet, methods to record

and check one's own form on video and to compare

one's own form with that of model players, such as

professionals using video-sharing services, are

widely used. In light of these current conditions, our

research aims to promote the utilization of knowledge

by digitizing amateur players' baseball notebooks and

to create a collaborative environment by sharing the

notebooks with others in the future. Furthermore, it is

desirable to be able to provide new insights that could

not be obtained only by the player's subjective view

by using AI to analyze videos of his or her form,

quantifying and visualizing his or her past and

differences from others in an easy-to-understand

manner. By providing these functions as a web

system that can be used only with a smartphone, we

aim to provide more growth opportunities to amateur

athletes who have not had access to state-of-the-art

equipment or a coaching environment.

Figure 1: The effect of our aimed training with digital

technology.

2.2 Body Tracking from Images

With the development of image analysis technology

using deep learning, it has become possible to detect

people from images with high accuracy. In recent

years, various skeletal points such as hands, feet, and

face can be recognized and output as coordinates.

Furthermore, several OSS libraries, such as

OpenPose (Qiao, 2017), have been released, making

it possible to easily implement skeletal recognition

functions from images into applications. Many

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

186

libraries support the output of 2D skeletal

coordinates, but libraries that estimate 3D coordinates

are also beginning to be provided. However, the

accuracy is currently low, and it is intended for

entertainment applications such as VR, so it is not

suitable for use as measurement data. In this study,

we applied trt_pose provided by Nvidia, considering

its licensing and ease of implementation, etc. The

coordinate points of the skeleton that can be

recognized by trt_pose are shown in Figure 2 and

Table 2. The trt_pose can recognize 18 joint points

defined in the MSCoCo dataset (Janardhanan, 2022).

In this study, since the baseball form is the target,

facial points are not necessary, and only the nose is

treated as the head, and the 14 points other than the

facial points are mainly used.

Table 1: Recognized joint points.

ID Joint name

ose

2 Right shoulde

3 Ri

ht elbow

4 Ri

ht wris

5 Left shoulde

6 Left elbow

7 Left wris

8 Right hip

9 Ri

ht knee

10 Ri

ht ankle

11 Left hip

12 Left knee

13 Left ankle

14 Right eye

15 Left e

16 Ri

ht ea

17 Left ea

Figure 2: Coordinates of joint points.

2.3 Motion Causal Network Analysis

Techniques

Using time series data of joint point coordinates

extracted from videos of pitching and hitting form, we

propose a motion network analysis technique to

visualize the linkage of motion between each skeletal

point in order to quantify and visualize the

characteristics of skeletal motion. This technique is

based on the technique for quantifying and

visualizing the linkage between players in group

sports. This technique uses transfer entropy (Staniek,

2008), quantifying the strength of the dependency

between two time-series data considering causality by

information theory. These include nonlinear

analytical methods that use mutual information

content, relative entropy, etc. Although these methods

are suitable for evaluating the strength of the

relationship between two signals, they are not suitable

for analyzing the direction of information flow, i.e.,

causality, because they are symmetrical with respect

to the two signals. In contrast, the application of

transfer entropy has recently been promoted as a

method suitable for causality estimation. It has been

used in research to estimate neurotransmission from

time series signals showing the activity of each

neuron in brain measurement, which is also a field of

human measurement (Wibral, 2014), and in research

such as communication analysis in human infants

(Hidaka, 2013), and it has been found that transfer

entropy can be applied even between signals with

nonlinear fluctuations. It has been found that the

transfer entropy can be applied between signals that

vary nonlinearly.

Figure 3 shows the process flow for calculating

the transfer entropy between two skeletal coordinates

and visualizing the entire skeletal network. First, in

process (1), skeletal coordinates are extracted from

the image of each frame of the video, time series data

of the coordinates of each skeletal point is generated,

and the amount of movement per second is calculated

from the time series data of the coordinates. In

process (2), the time-series data of the variable

(amount of movement) for each skeletal point is

transformed into a discrete random variable that is

normalized based on the histogram of the intensity of

movement of each skeletal point. In other words, they

are transformed into discrete states discretized by the

intensity of the movement. The number of divisions k

of the histogram was determined based on the Sturges

formula. l is the number of data in the time series data

of the variable. The amount of information transfer

between the variables of the two players is calculated

from the transfer entropy in the process (3). Here, if

Motion Causal Network Analysis for Quantitative Evaluation of Baseball Form by Video Analysis

187

the elements of the random variables I and J at time

step n are

𝑖



and 𝑗



, the transfer entropy T(J

→

I),

which indicates the influence of J on I, is calculated

by the formula in Figure 3, where

𝑃



𝑖



,𝑖



,𝑗





the simultaneous probability of

𝑖



, 𝑖



and 𝑗



, and

𝑃



𝑖



𝑖





denotes the conditional probability of

being

𝑖



when 𝑖



. Given two time-series data, I

and J, we can measure the degree of uncertainty that

changes relative to the uncertainty of predicting J's

next state from I's past series when we add J's past

series The transfer entropy is a measure of the degree

of uncertainty in the future state of J. The transfer

entropy takes values between 0 and 1; with larger

values, the more causally related I and J are. In this

study, since the time series data of joint point

coordinates extracted from the video was used, the

effect of the amount of movement of J on the amount

of movement of I in the next frame was calculated.

Figure 3: Processing flow of transfer entropy calculation.

In order to draw a network from the calculated

transfer entropy, the transfer entropy of all the paths

of two points from all the skeletal points is calculated,

and a directed graph is drawn by considering that

there is a linkage above a certain threshold value. The

number of graph occurrences can be adjusted by

adjusting the threshold value. The directed graph data

can be subjected to various network analyses, such as

centrality and page rank, allowing quantitative

analysis of network characteristics.

Figure 4: Visualization of causal network analysis.

Table 2: Definition of metadata.

Input item Contents

Type of ball

Hardball/ softball/ urethane/

sand/ other

Type of bat Material, length, weight

Condition of

the ground

Soil/ grass/ concrete/ mats

Type of shoes Athletic shoes/ spikes/ others

Weather

Sunny/cloudy/rainy/snowy,

temperature, wind speed

(tailwind/ head wind/ cross

wind)

Date and time yyyy/MM/dd HH:mm:ss

Comment free

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

188

2.4 Development of Prototype System

As a prototype system for this study, we constructed

a web system that allows users to upload videos of

their forms from their smartphones and visualize the

output results of motion network analysis. When a

user analyzes his/her form, he/she first prepares a

video of his/her form. If the form is a hitting form, the

user prepares a video of only one swing, trimmed in

advance using a smartphone function or the like.

When uploading the video, the conditions under

which the form was taken can be entered as metadata.

The input items are defined as shown in Table 3,

referring to the items noted in a typical baseball

notebook. The analysis results of each video can be

retrieved by using each item of metadata as a key, so

it is possible to compare the differences in form

between videos under various conditions or to

compare the changes before and after practice by

specifying the date and time. Figure 4 shows an

example of the results of applying motion network

analysis. In addition to the coordinates based on

skeletal recognition, the prototype system we

developed estimates the position of the center of

gravity, which is generally considered important in

form evaluation, and adds it to the joint point

coordinates. Furthermore, the tip of the bat used in

baseball hitting can be recognized by image

processing and added as one of the skeletal

coordinates, enabling evaluation of the direct effect of

body usage on the swing.

3 EVALUATION

To evaluate the validity and usefulness of the results

of motion network analysis using the proto-system,

we input videos of actual baseball hitting practice

forms and compared the results of transfer entropy

calculation with subjective evaluation. The subject

was a player belonging to a baseball team for working

people. 10 videos of each of the two patterns of hitting

(A and B) shown in Figure 5 were filmed and

analyzed. Pattern B is the hitting form of tee-batting

with a fixed ball, in which the subject player works

on improving his hitting form. We interviewed the

players about the points of improvement they were

aware of in Patterns A and B, and summarized the

correspondence with the hypotheses quantitatively

evaluated by motion network analysis in Table 4.

Improvement point (1) is related to the difference

between hitting a tossed ball and hitting a placed ball.

In placed tee batting, the player hits a fixed ball,

which stabilizes the form. Therefore, B is considered

to have less variation than A in the transfer entropy of

the motion network analysis. Improvement point (2)

is that in A, the player tries to hit the ball far away and

puts unnecessary force on the lower body, whereas in

B, the player is conscious of the stability of the lower

body. Therefore, it is thought that the variation of the

lower body value and the influence on other parts of

the body are reduced in the transfer entropy. Point (3)

is the use of the elbow and wrist in the upper body. in

order to improve the situation in A, where the bat

swing lags behind the body rotation (commonly

called the state where the upper body is open), in B,

the right elbow to the armpit is slightly opened, and

the upper body is rotated together with the left

shoulder to left elbow rotation, which results in a

more delayed right elbow and wrist rotation. As a

result, he is more conscious of delaying the external

rotation of the right elbow and wrist and rotating the

upper body (reducing the openness of the upper

body). Therefore, in the transfer entropy, there is a

linkage from the left elbow to the right elbow and

wrist. We quantitatively and statistically evaluated

whether the results of the movement network analysis

were valid for the hypotheses at these points of

awareness.

Table 3: Key points of improvement and hypothesis of data

analysis.

No.

Key Points of

Improvemen

Hypothesis of data

anal

sis

B is more stable

because the ball is

fixed.

The variation of

transfer entropy in

B is smaller than in

B is more aware

of lower body

stability.

The transfer

entropy starting

from some lower

body point

decreases.

In B, the right

elbow to the side

is slightly open,

and the right

elbow and wrist

are delayed in

rotation.

Transfer entropy

from the left elbow

to the right elbow

and wrist tends to

be higher.

Motion Causal Network Analysis for Quantitative Evaluation of Baseball Form by Video Analysis

189

Figure 5: Sequential images of the evaluated batting forms.

(A) Toss batting with normal awareness (B) Tee batting

with awareness of points for improvement.

* p<0.05 (Welch’s t-test)

Figure 6: Mean of the transfer entropy starting at each joint.

Figure 7: Mean of the transfer entropy toward each joint.

4 RESULT

4.1 Improvement Points (1) and (2)

To evaluate improvement point (1), we compiled the

transfer entropy for each swing of A and B and

compared the mean value of transfer entropy where

each site is the starting point with the mean value of

transfer entropy where each site is the ending point,

as shown in Figures 6 and 7. To test the hypothesis of

improvement point (1), Table 5 summarizes the

standard deviations of the values for each site, and as

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

190

shown in Figures 6 and 7, A has many outliers and a

large variation in values. The standard deviation

values in Table 5 are also significantly lower for B for

the transfer entropy at the starting point. On the other

hand, a decreasing trend was observed for B for the

endpoint transfer entropy, although it was not

significant. In other words, the results are in

accordance with the hypothesis that B has less

variation in the evaluation using the transfer entropy.

To evaluate the improvement point (2), a t-test

(Welch's t-test) was performed for each site to

compare the distribution of the transfer entropy of the

starting point of each site, as shown in Figure 6. The

significance level was set at 0.05. The results showed

that the mean value of B, which is the starting point,

decreased in the left and right hip, knee, and ankle,

which are the lower body regions, and that there was

a significant decrease, especially in the right_hip,

left_hip, and right_ankle. This result is in accordance

with the hypothesis that the lower body's extra

movement decreases due to the awareness of lower

body stability and that the transfer entropy as the

starting point decreases. In addition, there was also a

significant decrease in the starting point of transfer

entropy at the nose, which is the head, indicating that

not only the lower half of the body but also the upper

half of the body showed a decrease in extra

movement.

* p<0.05 (Welch’s t-test)

Figure 8: Mean of transfer entropy from left_elbow to each

joint.

4.2 Improvement Points (3)

To evaluate the improvement point (3), we compared

the transfer entropy for each site starting from

left_elbow as shown in Figure 8. The results showed

that the transfer entropy increased significantly in B

for the left_wrist, right_wrist, and right_elbow, which

are related to the improvement point (3). As a result,

less than 5 out of 10 swings were calculated correctly,

which was insufficient for statistical evaluation. In

addition, there was a significant increase in transfer

entropy for the left_knee and a significant decrease

for the right_ankle, which were not included in the

hypothesis for improvement point (3). Although these

results were not included in the players' awareness,

improvement points (2) and (3) resulted in increased

rotation of the upper body and knee regions.

5 DISCUSSION

The validity of the results of this report was examined

based on interviews with the players and teams

included in this study. In the motion network analysis

of form A→B, which the players consciously aimed

to improve, reasonable results close to the hypothesis

were obtained for improvement points (1), (2), and

(3). For improvement point (1), outliers were reduced

at B in the transfer entropy of the start and end points,

and the variability was reduced. On the other hand,

the significant decrease only at the starting point may

be because the end point is less stable than the starting

point, which is a characteristic of the subject athletes.

In the improvement point (2), a significant decrease

was observed, especially in the hip near the waist,

which may be because the knee has large movement

and is prone to recognition errors. Further verification

of the difference in awareness of the knees and hips

in the form is needed. In the improvement point (3),

the linkage between the left_elbow and other wrists

and elbows changed as hypothesized. However, the

skeletal structure of the right hemisphere, which is

hidden during the swing, was difficult to recognize

from the photographer's side, and this did not lead to

sufficient verification. In this respect, improvements

are needed, such as in the shooting method and in the

use of skeletal recognition libraries that are relatively

accurate even for the hidden parts.

The usefulness of this technology was also

examined in the same way. We presented the linkage

of each part visualized in the motion network analysis

to the players and obtained an evaluation that it was

intuitively easy to understand. This is useful for

confirming the points that players were conscious of

in practice, including the validity of the hypothesis.

In addition, the metadata collected simultaneously

was also evaluated as having the necessary

information for form evaluation based on interviews

Motion Causal Network Analysis for Quantitative Evaluation of Baseball Form by Video Analysis

191

with players and teams. On the other hand, there are

many points in the motion network analysis where it

is difficult to understand the network formation by

transfer entropy, except in the hypothesized areas. In

the future, it will be necessary to collect more data on

more forms and to increase the number of patterns in

interpreting the meaning of the starting and ending

points of the transfer entropy through repeated

verification of the hypothesis. Since only two patterns

of forms were tested in this report, we did not go as

far as to compare data from different environments

utilizing metadata or to compare data from other

people. Methods and designs to make players aware

of the ability to accumulate and share large amounts

of data also need to be considered.:

6 CONCLUSIONS

We proposed a motion network analysis method that

recognizes the positional coordinates of the skeletal

structure from videos of athletes taken with

smartphones, quantifies the linkage of the movements

of each part using transfer entropy, and visualizes the

relationship between them to promote effective

training methods that utilize data in the amateur

sports industry. Furthermore, we developed a

prototype system that visualizes characteristic points

from videos of baseball forms by motion network

analysis and records and stores notes on the

environment and situation at the time of filming. In

order to evaluate the proposed method and the

prototype system, we analyzed videos of one amateur

baseball player, taking 10 videos each of tee batting

with a normal front toss and the form of a replacement

tee batting that he worked on with an awareness of

improvement, and formulated a hypothesis for the

data regarding the three points of awareness and

conducted a statistical evaluation. Statistical

evaluation was conducted. The results of the analysis

showed significant changes (p<0.05) consistent with

the hypotheses, including a reduction in the

variability of transfer entropy due to stabilization of

form and an improvement in wrist and elbow

coordination due to changes in the use of the right

elbow. This evaluation confirmed that the players

themselves intuitively understood the effects and

noticed the points that they worked to improve. Based

on these results, we conclude that the proposed

method can be used to easily identify the

characteristics of players' form and may contribute to

improving the efficiency of training for a larger

number of players through the accumulation of data

and hypothesis testing.

REFERENCES

Chung, J. L., Ong, L. Y., & Leow, M. C. (2022).

Comparative analysis of skeleton-based human pose

estimation. Future Internet, 14(12), 380.

Qiao, S., Wang, Y., & Li, J. (2017). Real-time human

gesture grading based on OpenPose. In 2017 10th

International Congress on Image and Signal

Processing, BioMedical Engineering and Informatics

(CISP-BMEI) (pp. 1-6). IEEE.

Itoda, K., Watanabe, N., & Takefuji, Y. (2015). Model-

based behavioral causality analysis of handball with

delayed transfer entropy. Procedia Computer Science,

71, 85-91.

Janardhanan, J., & Umamaheswari, S. (2022, March). A

comprehensive study on human pose estimation. In

2022 8th International Conference on Advanced

Computing and Communication Systems (ICACCS)

(Vol. 1, pp. 535-541). IEEE.

Staniek, M., & Lehnertz, K. (2008). Symbolic transfer

entropy. Physical review letters, 100(15), 158101.

Wibral, M., Vicente, R., & Lindner, M. (2014). Transfer

entropy in neuroscience. Directed information

measures in neuroscience, 3-36.

Hidaka, K. (2013). Visualization of Information Network in

Human Multimodal Communication. The Journal

of the Institute of Electronics, Information and

Communication Engineers, 96(12), 945-950.

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

192