Detection of Shot Information Using Footwork Trajectory and

Skeletal Information of Badminton Players

Naoki Tanaka

, Hidehiko Shishido

, Masashi Suita

, Takeshi Nishijima

, Yoshinari Kameda

and Itaru Kitahara

Master’s and Doctoral Program in Intelligent Mechanical Interaction Systems, University of Tsukuba, Ibaraki, Japan

Center for Computational Sciences, University of Tsukuba, Ibaraki, Japan

Faculty of Health and Sport Sciences, University of Tsukuba, Ibaraki, Japan

Faculty of Health Sciences, Tokyo Metropolitan University, Tokyo, Japan

suita.masashi.gn@u.tsukuba.ac.jp, t-nishijima@tmu.ac.jp

Keywords: Shot Information, Footwork Trajectory, Skeletal Information, Video Analysis, Badminton.

Abstract: As video analysis has become important for sports science, various research has been conducted. In badminton,

while shot information is essential primary data for performance analysis, it has been input manually, which

makes it difficult to give instant feedback onsite. Our research aims to automatically detect shot information

from videos of badminton game. By applying video tracking, the player's footwork trajectory and skeletal

information are estimated. Based on the estimated information, the hit timing is detected using deep learning

classification. The horizontal position of the hit point, which is useful for game analysis, is also detected from

the player's footwork trajectory around the hit timing.

1 INTRODUCTION

With the development of IT, data analytics is being

considered fundamental for sports science, especially

in competitive sports (Pascual, 2006), (Messelodi,

2019). Since analysis using video/image data is par-

ticularly noteworthy as information can be confirmed

visually, there are active research to analyze players'

body information in badminton competition (Kokum,

2017), (David, 2014) and shot information (Chen,

2007), (Nyan, 2020), (Yoshikawa, 2021). These re-

search detect players' forms and tactics to improve

their competitive performance.

A badminton game consists of multiple rallies in

which players exchange shots, and each rally includes

a variety of shots. Thus, the shot information is con-

sidered as the primary data for analysing the evolution

of rallies that lead to scored and lost. On this demand,

analytical software such as “Sportscode” is widely

https://orcid.org/0000-0001-8575-0617

https://orcid.org/0000-0002-6765-4282

https://orcid.org/0000-0001-6776-1267

https://orcid.org/0000-0002-5186-789X

used for video analysis of badminton shot infor-

mation. It can analyze player performance by creating

a database of competition scenes. For data aggrega-

tion, the timing of shots taken by players and patterns

of goals scored is manually input. On the other hand,

badminton matches take more than one hour, and col-

lecting data using manual input requires lots of time

and effort.

In order to solve the problem of badminton game

analysis, as shown in Figure 1, this research proposes

a method for automatically detecting shot information

from monocular badminton singles match videos. The

detected shot information consists of two properties

of the hit point: “Hit timing” and “Horizontal posi-

tion".

• [Hit timing] The time when the player hits a

shuttle.

• [Horizontal position] Vertical projection of the

3D position where a racket hits a shuttle onto the

court (ground).

112

Tanaka, N., Shishido, H., Suita, M., Nishijima, T., Kameda, Y. and Kitahara, I.

Detection of Shot Information Using Footwork Trajectory and Skeletal Information of Badminton Players.

DOI: 10.5220/0012162700003587

In Proceedings of the 11th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2023), pages 112-119

ISBN: 978-989-758-673-6; ISSN: 2184-3201

Figure 1: Shot information detection and tactical analysis using footwork trajectory and skeletal information. The shot infor-

mation is detected using footwork trajectory (upper-left) and skeletal information (lower-left) from a monocular video of a

badminton game. The detected shot information consists of the hit timing and the horizontal position of each hit point (center).

Tactics are analyzed by visualizing the ratio of shot direction based on the detected shot information as shown in the right.

In this research, it is assumed that the positions of

the players and their dominant wrists show time-se-

ries changes correlation with the hit timing. Since

these time-series changes are more pronounced in sin-

gles matches, this research focuses on singles matches.

Further contributes of this research is to solve the data

collection problem by evaluating tactical analysis us-

ing the detected data.

2 RELATED WORKS

2.1 Human Pose Estimation

Some 3D tracking sensors such as “Vicon” and

“OptiTrack” require attaching markers to the body for

estimating a subject's pose. Although, this kind of

method has high accuracy in pose estimation (the

skeletal position with an error of 0.1 mm or less), ap-

plying this method to a live sporting event is difficult,

due to some limitation to the players' movement,

Some methods can estimate a subject's pose from

images and videos (Zhe, 2019), (Sun, 2019). There

are two major approaches to this method: top-down

and bottom-up. The top-down method first detects

subjects using object detection and then detects joints

for each person. On the other hand, the bottom-up

method detects all joint points in an image without

object detection and then compiles the joint points

constituting each subject. While the top-down method

is generally more accurate, the processing time in-

creases significantly when the number of subjects in-

creases. On the other hand, the disadvantage men-

tioned above is less likely to occur due to the small

and fixed number of players in the badminton game

video used in this research. Therefore, a top-down

person poses estimation method is used in this re-

search.

Human pose estimation from images has been ap-

plied to various kind of sports. For example, Dejan et

al. proposed a system to estimate the pose of a jumper

and a skiboard in ski jumping videos and evaluate the

pose of the jumper during jumping (Dejan, 2022). No-

naka et al. proposed a system to estimate the pose of

a player in a tackle scene in a rugby video and to judge

dangerous tackles that may cause concussions (No-

naka, 2022). Kaustubh et al. estimated players' poses

in table tennis videos (Kaustubh, 2021). They classi-

fied hitting styles using deep learning with the time-

series information as input, achieving more than 99%

accuracy in 11 hitting style classification.

2.2 Badminton Shot Information

Analysis

Shot information is helpful for tactical analysis of

badminton games. Wei et al. classified shots into six

types (five plus others) based on players' pose infor-

mation at hit timing (Wei, 2017). Wei et al. also ana-

lyzed the game situation based on the usage of each

shot. The analysis results were compared with the ac-

tual match results. It was confirmed that the results of

the game situation analysis obtained from the shot in-

formation were equivalent to the actual match results.

In this research, we attempted to detect shot infor-

mation in consideration of its applicability to the tac-

tical analysis of badminton.

2.3 Detection of Shot Information

Shuttle tracking has been reported as one of the most

Detection of Shot Information Using Footwork Trajectory and Skeletal Information of Badminton Players

113

common shot information detection methods. By

tracking the shuttle, information on the shuttle's posi-

tion and trajectory can be obtained and information on

each shot can be detected. There is an example of de-

tecting the shuttle in an image using the image sub-

traction method and comparing the shuttle's trajectory

with various strokes (Chen, 2007). In another case

(Nyan, 2020), the authors applied Tracknet and poly-

nomial curve fitting to the shuttle's trajectory using

deep learning and attempted to detect shot infor-

mation. All these research assume that the shuttle is

always observed within the angle of view. However,

since badminton competition video is generally shot

from a single fixed camera focusing on the players,

there are frames in which the shuttle is not seen in the

video. These frames may make it difficult to track the

shuttle.

Therefore, a method to detect shot information

without tracking shuttles was proposed (Yoshikawa,

2021). The hit timing is identified from the player's

skeletal information. The shot direction is detected by

connecting the hit points of the two players at the hit

timing in a time series. However, since there are

countless possible forms in badminton depending on

the shuttle's position, more than classification based

on a small number of conditional branches is required

to identify the hit timing for all forms.

This research aims to detect shot information us-

ing players' footwork trajectories and skeletal infor-

mation. By achieving this objective, the following

two contributions can be obtained.

•

The detection of shot information that can be ap-

plied to match videos, including scenes in which

the shuttle is not visible.

•

The detection of shot information that can be ap-

plied to various shots.

3 PRELIMINARY EXPERIMENTS

Our research uses a player's 2D skeletal information

(16 - 25 joint points) from monocular videos and shot

information detection. However, there are some prob-

lems with inputting all the acquired skeletal infor-

mation into the deep learning model as follows.

• The accuracy of shot information detection is

highly dependent on the accuracy of the conven-

tional human skeleton estimation method.

• Increased computational complexity and over-fit-

ting risk due to many input dimensions.

Figure 2: Relationship between time-series changes in player position (upper) and wrist position (lower), respectively, and

hit timing

Figure 3: Overview of the shot information detection method. The skeletal information of the two players is detected from

the monocular video, and the player's position and dominant wrist position are extracted. These are input to LSTM to detect

the hit timing and the ground projection of the hitting point. “additional detection” is explained in Section 4.2.

icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support

114

Thus, in this research, only the player's position (the

midpoints of both ankles) and wrist position (the rela-

tive position of the dominant wrist from the waist) are

used among the acquired skeletal information to ad-

dress the abovementioned problem. These features

were selected in consideration of the features of bad-

minton games, and their validity as features were veri-

fied by visualizing the relationship with the hit timing.

In badminton singles, the base is usually located

near the center of the court. The player has the feature

of moving back to the base after his shot to prepare

for the opponent's hit. Therefore, as shown in Figure

2 (left), the time-series variation of the player's posi-

tion (x, y) is correlated with the hit timing. Further-

more, since the player's position at the hit timing is

the position where the player moved to release the

shot. Thus, the player's position around the hit timing

and the horizontal position of the hit point are close to

each other.

The player's dominant hand holding the racket has

the postural features of being the most distant from

the body center at the hit timing. Therefore, as shown

in Figure 2 (right), the relative position of the domi-

nant wrist from the waist (x, y) also correlates with

the hit timing. Therefore, this information has features

time-series changes at the hit timing and practical fea-

tures for detecting the hit timing and the horizontal

position of the hit point.

4 DETECTION OF SHOT

INFORMATION

As shown in Figure 3, this research proposes a method

for detecting shot information, such as hit timing and

the horizontal position of the hit point. A deep learn-

ing based approach is taken where the network takes

the physical features at the hit timing as the input. As

shown in Figure 4, the conventional method of human

skeleton estimation (Sun, 2019) is applied to a bad-

minton competition video to obtain the 2D skeletal in-

formation of the player. The player's and wrist posi-

tions are calculated based on the skeletal information.

Hit timing is detected from the features of the time

series data of the player's and wrist positions. The hor-

izontal position of the hit point can be detected from

the player's footwork trajectory centered on the de-

tected hit timing.

4.1 Detection of Hit Timing

This section describes the preprocessing and output

data for the input data of the hit timing (as shown in

Figure 1: f(i)) detection model.

As shown in Figure 5, the elements used as input

data for the model are the time-series data of player

positions and wrist positions obtained in Section 4.1.

The player position data is for two players, one who

releases the shot and the other who receives the shot.

The wrist position data is for one player who releases

the shot. A sliding window of step s and length w is

applied to these time series data to generate partial

time series data. The model uses LSTM with the input

partial time-series data at each time point. The model

output the probability that a hit timing is included in

the time of the partial input time series data.

Figure 4: Acquisition of skeletal information and player po-

sition. Bule crosses is the midpoint of both ankles and de-

fined as the player position.

Figure 5: Process flow of hit timing detection method. A sliding window is applied to the time series data of the positions of

the two players and their dominant wrists to generate partial time series data. The generated partial time-series data is input

to LSTM for hit timing detection. "Conversion to frame number" and "additional detection" are explained in section 4.2. d:

dimensionality of features, w: window width of a sliding window, n: number of partial time series.

Detection of Shot Information Using Footwork Trajectory and Skeletal Information of Badminton Players

115

Cross entropy error between the correct label and

the output probability is calculated. The model is

trained to make the error small.

Furthermore, the following condition is applied to

the hit timing detection results of the two players to

reduce the number of undetected hit timing.

• If the hit timing of one player is detected con-

secutively, the frame with the largest output

probability for the other player between the two

consecutive hit timing is additionally detected as

a hit timing.

This condition is to take advantage of the competitive

features that two players always alternate shots in a

badminton game.

The accuracy of hit timing detection can be eval-

uated by comparing the final hit timing prediction

frame (Figure 5: hit timing') with the ground truth af-

ter adding this additional detection.

4.2 Detection of the Horizontal Position

of the Hit Point

This section describes the preprocessing for the mod-

el's input and output data for detecting the horizontal

position of the hit point (as shown in Figure 1: (x(i),

y(i))).

As shown in Figure 6, the elements used as inputs

to the model are the time series data of player position

obtained in Section 4.1 and the hit timing obtained in

Section 4.2. From the time series data of player posi-

tions, partial time series data of length w centered at

each hit timing is extracted. Using LSTM with the

partial time series data of each hit timing as input, the

model output the 2D coordinates of the horizontal po-

sition of the hit point.

MSE (mean square error) is calculated using the

ground truth as the horizontal position of the hit point.

The model is trained so that the error is small.

5 EXPERIMENTS

The detection accuracy of the hit timing and the hori-

zontal position of the hit point was verified, respec-

tively. Furthermore, from the analysis results using

the shot directions obtained from the horizontal posi-

tion of the hit points of two players, the possibility of

using the detected data for analysis is confirmed.

5.1 Experiment Setup

The data used in the validation are videos of two

games (four sets in total) in Akane Yamaguchi vs. Tai

Tzu Ying in 2019 and 2021, with each set as Video

No.1 to 4. Videos No. 1 to 3 were used as training data

and videos No. 4 as test data. HRNet (Sun, 2019) was

used to extract the skeleton of the players. HRNet is

an algorithm that estimates the skeletal shape of a per-

son in a 2D image by deep learning and is capable of

real-time detection even when multiple people are

mixed in the image. The parameters in the algorithm

are s=5, w=40 and df=14. Precision, recall, and f-

measure are used as hit timing detection accuracy in-

dices.

5.2 Experiment to Detect the Hit Timing

In this research, the detected hit timing is used to de-

tect the horizontal position of the hit point. However,

the impact of an error of a few frames on the hit timing

detection is small. Because the model inputs the foot-

work trajectories of players for several tens of frames

around the hit timing. Therefore, this research does

not require strict accuracy of hit timing frame detec-

tion in units of a few frames. Considering the foot-

work speed of the player, a specific error (E frames)

from the ground truth of the hit timing frame can be

tolerated. In this experiment, E=20 (about 0.67 sec-

onds) was set based on observing player speed in the

Figure 6: Processing flow of the method for detecting the horizontal position of the hit point. Extract the time series data of

the player's position around the hit timing. The extracted time-series data is input to LSTM to detect the horizontal position

of the hit point. d': dimensionality of features, w: window width of a sliding window, h: number of hit timing.

icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support

116

Table 1: Accuracy evaluation of hit timing detection.

data. In other words, a detection whose error from the

ground truth of the hit timing frame is within 20

frames is defined as a correct detection.

Table 1 shows the hit timing detection accuracy

for the players at the front of the court. The three types

of data used for hit timing detection are denoted

as pos

(position of the player at the front of the

court), pos

(position of the player at the back of the

court), and wri

(wrist position of the player at the

front of the court). The detection accuracy of the

model when these data are combined and input, re-

spectively, is shown. The model with the additional

detection described in section 4.2 is denoted as +add.

5.3 Experiments to Detect the

Horizontal Position of the Hit Point

Since hit timing is used to detect the horizontal posi-

tion of a hit point, the accuracy of hit timing detection

affects the accuracy of the detection of the horizontal

position of the hit point. In this section, the ground

truth is used for the hit timing to evaluate only the de-

tection of the horizontal position of the hit point. As

shown in Figure 7, the x-axis is set parallel to the net,

and the y-axis is set perpendicular to the net, with the

yellow circle in the upper left corner as the origin in

the overhead view of the court. Considering the half-

court size of an actual badminton singles court

(5.18[m]×6.70[m]), the output coordinates are set

as x=[0~5.18] and y=[0~6.70]. RMSE (root mean

square error) is used to measure detection accu-

racy. RMSE of the output and the ground truth for all

test data was RMSE=0.54[m].

Figure 7: Coordinate system of the projected position of the

hit point on the ground.

5.4 Evaluation of Analysis by

Percentage of Shot Direction

This section describes the tactical analysis of players

using the horizontal position of the hit points detected

by this method and the possibility of applying the de-

tected data to the analysis. As an example of analysis,

this section presents an analysis method that divides a

half-court into nine (3×3) areas (Josue, 2020), (Ca-

reelmont, 2013). This analysis reveals player features

based on the percentage of shot directions from a spe-

cific area. Among the nine areas of the court surface,

the hit area is defined as the area that includes the hor-

izontal position of the hit point, and the shot direction

can be detected by connecting the hit areas (or the hor-

izontal position of the hit point) of two players in a

time series. The ratio of shot directions can be used

for tactical analysis as an indicator of the features of

each player's game.

Using the correct label and detected hit timing

data used in this method, the horizontal position of the

hit point can be detected from the three models shown

in Table 2. The G-model is a model that uses the

ground truth of the horizontal position of the hit point.

The GD-model is a model that uses the horizontal po-

sition of the hit point detected using the ground truth

of the hit timing. The DD-model is a model that uses

the horizontal position of the hit point detected using

the detected hit timing. Therefore, the three models

can also detect the hit area and shot direction.

The features of players can be visualized by plot-

ting the distribution of shot direction as a heat map.

Furthermore, by comparing the heat maps for each in-

put information, the applicability to the detection data

analysis can be qualitatively evaluated. Figure 8

shows the heat maps of the distribution of shot direc-

tions in the hit area where the number of hits by play-

ers in the front of the court was high in the test data.

The hit timing and the horizontal position of the hit

points detected by this method can lead to the same

level of analysis results as those using the ground

truth.

input recall(%) precision(%) f_measure(%)

79.3 82.1 80.7

76.0 79.3 77.6

91.7 94.1 92.9

80.2 78.9 79.5

90.1 91.6 90.8

95.9 94.3 95.1

93.4 96.6 95.0

96.7 95.1 95.9

𝑝𝑜𝑠

𝑓

𝑝𝑜𝑠

𝑏

𝑤𝑟𝑖

𝑓

𝑝𝑜𝑠

𝑓

, 𝑝𝑜𝑠

𝑏

𝑝𝑜𝑠

𝑓

, 𝑤𝑟𝑖

𝑓

𝑝𝑜𝑠

𝑏

, 𝑤𝑟𝑖

𝑓

𝑝𝑜𝑠

𝑓

, 𝑝𝑜𝑠

𝑏

, 𝑤𝑟𝑖

𝑓

𝑝𝑜𝑠

𝑓

, 𝑝𝑜𝑠

𝑏

, 𝑤𝑟𝑖

𝑓

𝑎𝑑𝑑

Detection of Shot Information Using Footwork Trajectory and Skeletal Information of Badminton Players

117

Table 2: Three models for detecting the horizontal position

of the hit point.

Figure 8: Visualization of shot direction percentage from

the right front area of the court. The shot directions detected

from the G-model (left), GD-model (center), and DD-model

(right) are visualized using heat maps. The yellowish region

indicates the most likely location of the endpoint of the shot

direction.

5.5 Limitations in this Research

This section describes the limitations of the detection

accuracy and the primary cause of error in this study.

It is difficult to detect hit timing in a single frame

using this research method. This is because the output

information (hit timing) is not uniquely determined

from the time-series changes in the input information

(movement trajectory and skeletal information).

Therefore, applying this method to tactical analysis is

difficult, which requires hit timing accuracy on a

frame-by-frame basis.

The primary cause of error in the experiment de-

scribed in Section 5.2 is the irregular movement of

players in hit timing. In this study, the following hy-

potheses are formulated for the movement trajectory

and skeletal information, respectively, as the regular

movements of the players.

• The player moves away from the base around

the hit timing and returns to the base again.

• At the hit timing, the player's wrist is farthest

away from the waist.

Therefore, detection errors are likely to occur when

players' movements that do not conform to the above

hypotheses (irregular movements) are observed.

We have attempted to reduce the number of errors

by applying the additional detection described in Sec-

tion 4.2, but we have not been able to address all er-

rors.

6 CONCLUSIONS

This research proposes a method for detecting shot in-

formation using badminton players' footwork trajec-

tories and skeletal information. The proposed method

detects the hit timing using the time series information

of the footwork trajectory and skeletal information,

which are characterized by the hit timing. Further-

more, this method detected the horizontal position of

the hit point using the footwork trajectory around the

hit timing. As a result of the demonstration experi-

ment, the hit timing detection accuracy (F-measure)

for the player at the front court was 95.9%, and the

horizontal position of the hit point was detected with

an accuracy of MSE=54.0. Using the detected data,

we quantitatively and qualitatively evaluated that the

tactical analysis can be performed at the same level as

the ground truth. As a future work, we plan to expand

the experimental data. We will confirm the generality

of the method from the results of tests using players'

data that are not included in the training data. We will

also investigate how players' playing styles and other

factors affect detection accuracy.

This method can automatically detect data for tac-

tical analysis using only the player's footwork trajec-

tory and skeletal information from the monocular

camera images. In conclusion, this research suc-

ceeded in solving the problem of the conventional

method using shuttle tracking and in automating the

data collection for badminton video analysis.

REFERENCES

Pascual J. Figueroa, Neucimer J. Leite, Ricardo M. L. Bar-

ros, (2006) “Tracking soccer players aiming their kine-

matical motion analysis, ” Computer Vision and Image

Understanding Vol.101, pp.122-135.

S. Messelodi, C.M.Modena, V.Ropele, S.Marcon, M.Sgro,

(2019) “A Low-Cost Computer Vision System for Real-

Time Tennis Analysis, ” ICIAP, pp.106-116.

Kokum Weeratunga, Anuja Dharmaratne, Khoo Boon

How, (2017) “Application of computer vision and vec-

tor space model for tactical movement classification in

badminton, ” IEEE Computer Society Conference on

Computer Vision and Pattern Recognition Workshops,

vol.2017, pp.132-138.

David Alder, Paul R.Ford, Joe Causer, A.Mark Williams,

(2014) “The coupling between gaze behavior and oppo-

nent kinematics during anticipation of badminton shots,

” Human Movement Science, Vol.37, pp.167- 179, Oct.

Chen Bingqi, Wang Zhiqiang, (2007) “A Statistical Method

for Analysis of Technical Data of a Badminton Match

Based om 2-D Seriate Images, ”Tsinghua Science and

Technology ISSN 1007-0214 14/19 pp.594-601,

Vol.12, No.5.

icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support

118

Nyan-Ping Ju, Dung-Ru Yu, Tsi-Ui Ik, Wen-Chin Peng,

(2020) “Trajectory-based Badminton Shots Detection, ”

ICPAI, pp.64-71.

Yui Yoshikawa, Hidehiko Shishido, Masashi Suita,

Yoshinari Kameda, Itaru Kitahara, (2021) “Shot detec-

tion using skeleton position in badminton videos, ”

SPIE Proceedings, Vol.11766, IWAIT.

Shayne Vial, Jodie Cochrane, Anthony J. Blazevich and

James L. Croft, (2019) “Using the trajectory of the shut-

tlecock as a measure of performance accuracy in the

badminton short serve,” International Journal of Sports

Science & Coaching, Vol.14, pp.91-96.

Rusdiana A. 3D Kinematics Analysis of Overhead Back-

hand and Forehand Smash Techniques in Badminton.

Ann Appl Sport Sci. 2021; 9 (3).

Zhe Cao, Tomas Simon, Shih-En Wei, Yaser Sheikh,

(2019) “Realtime Multi-Person 2D Pose Estimation us-

ing Part Affinity Fields,” IEEE Transactions on Pattern

Analysis and Machine Intelligence, 14 pages.

K. Sun, B. Xiao, D. Liu, J. Wang, (2019) "Deep High-

Resolution Representation Learning for Human Pose

Estimation, " CVPR, pp.5693-5703.

Dejan Stepec, Danijel Skocaj, (2022) “Video-Based Ski

Jump Style Scoring from Pose Trajectory,” IEEE/CVF

Winter Conference on Applications of Computer Vision

Workshops, pp.682-690.

Naoki Nonaka, Ryo Fujiwara, Monami Nishio, Hidetaka

Murakami, Takuya Tajima, Mutsuo Yamada, Akira

Maeda, Jun Seita, (2022) “End-to-End High-Risk

Tackle Detection System for Rugby,” IEEE/CVF Con-

ference on Computer Vision and Pattern Recognition

Workshops, pp.3549-3558.

Kaustubh Milind Kulkarni, Sucheth Shenoy, (2021) “Table

Tennis Stroke Recognition Using Two-Dimensional

Human Pose Estimation,” IEEE/CVF Conference on

Computer Vision and Pattern Recognition Workshops,

pp.4571-4579.

Wei-Ta Chu, Samuel Situmeang, (2017) “Badminton Video

Analysis based on Spatiotemporal and Stroke Features,”

ICMR, pp.448-451.

F. Josue, M. F. Abdullah, I. Zulkapri et al., (2020) “Move-

ment pattern in term of court coverage among top inter-

national male and female badminton players during

BWF World Championships 2013,” Jurnal Sains Sukan

& Pendidikan Jasmani, vol. 9, no. 1, pp. 9–14.

Careelmont S, (2013) “Badminton shot classification in

compressed video with baseline angled camera,” Mas-

ter Thesis, University of Ghent.

Detection of Shot Information Using Footwork Trajectory and Skeletal Information of Badminton Players

119