Ttswing: a dataset for table tennis swing and racket kinematics analysis

ABSTRACT We introduce TTSwing, a novel dataset designed to analyze table tennis swings. The dataset was collected using custom racket grips embedded with 9-axis motion sensors, which provide

precise kinematic data on swings. In addition, we provide anonymized demographic data for players. The dataset was collected from 93 participants, all of whom are elite table tennis players

from Taiwan. We detail the data collection and annotation procedures. These data are expected to improve the understanding of player performance and facilitate the development of tailored

training programs and biomechanical analyses, offering practical benefits to both athletes and coaches. TTSwing has excellent potential to facilitate innovative research in table tennis

analysis and is a valuable resource for the scientific community. We release the dataset and the experimental codes for reproducibility. SIMILAR CONTENT BEING VIEWED BY OTHERS OPTICAL MOTION

CAPTURE DATASET OF SELECTED TECHNIQUES IN BEGINNER AND ADVANCED KYOKUSHIN KARATE ATHLETES Article Open access 18 January 2021 SUMMARIZING PHYSICAL PERFORMANCE IN PROFESSIONAL SOCCER:

DEVELOPMENT OF A NEW COMPOSITE INDEX Article Open access 24 June 2024 A LARGE-SCALE MULTIVARIATE SOCCER ATHLETE HEALTH, PERFORMANCE, AND POSITION MONITORING DATASET Article Open access 30

May 2024 BACKGROUND & SUMMARY Since its inclusion as an Olympic sport in 1988, table tennis has gained widespread popularity. It is enjoyed around the world as a competitive sport and a

common recreational pastime among players of all levels and ages. Meanwhile, with advances in sensor technologies and machine learning algorithms, there is increasing interest in using

data-driven approaches to analyze sports, including table tennis. Such approaches provide valuable insight into player performance and inform training programs for players. This paper

introduces the TTSwing (Table Tennis Swing) dataset, a novel dataset that includes swing information collected by the 9-axis sensors embedded in the grips of customized paddles. The swing

here refers to the powerful forehand smash movement. Since the sensor is embedded in the handler of a racket, perhaps a more precise term for swing information is racket kinematics. We use

these two terms interchangeably in the following. TTSwing dataset accurately details racket kinematics, offering critical insights into player performance, shot quality, and technical

differentiation among skill levels1,2,3. In addition to swing information, the dataset includes anonymized demographic details of players, such as age, gender, height, weight, racket-holding

hand, and years of experience in the game. Combined with kinematic data, these demographic attributes allow for an in-depth analysis of the interplay between player characteristics and

performance metrics, enabling the development of customized training programs and biomechanical studies. This comprehensive dataset provides a valuable resource for studying table tennis

swings and can be used to develop new techniques and approaches to improve table tennis skills. Previous studies have taken advantage of advances in sensor technology to collect detailed

information on athletes or user movements4,5,6. For example, researchers attached a DELSYS sensor to ten points in the right arm to collect muscle information7. They also analyzed the

differences between professional and amateur players when stroking. A follow-up paper expanded the collected data by placing a hang3.0 sensor in nine different areas, e.g., hands, limbs, and

waists8. The authors classified motions according to acceleration and angular velocity and used the results to improve the stroke posture of the players. Some studies have placed the smart

device on the wrist of the player to collect acceleration and angular velocity data during stroke9,10,11. However, using excessive portable devices may make player movements less natural and

may indirectly impact the player’s performance. Another work embedded sensors in the grip of the racket12 and predicted the spin of table tennis and stroke type13,14. Compared to other

collection methods, embedding sensors into the grip reduces the burden on the player and makes data collection easier. In addition, the embedded sensor in the grip captures the racket

movement, allowing the detection of even subtle changes in racket movement. In our paper, we have chosen to utilize the last method for data collection. However, previous studies only

collected swing information from a limited number of players, typically ranging from a few to a dozen12,13,14. In contrast, our dataset is a unique resource for future research, as we

collected and analyzed information from nearly 100 elite players, providing a much larger sample size than in previous experiments. In summary, our work has the following contributions. * We

present a new dataset, TTSwing, that captures professional table tennis players’ swings (racket kinematics) along with anonymized demographic information of players. We describe the

collection and annotation process. To our knowledge, this is the largest open dataset for professional players’ swing information with their anonymized demographic information12– 14. * The

dataset provides a high-resolution record of racket movement, enabling future research in table tennis analytics, biomechanics, and player performance assessment. By offering structured and

well-annotated swing data, TTSwing can serve as a valuable resource for studies on skill evaluation, training optimization, and data-driven sports science applications. * To support

reproducibility, we openly release the dataset together with structured documentation detailing the data attributes, the collection process, and potential use cases. This dataset lays a

foundation for future research in table tennis swing analysis and broader applications in sports technology. METHODS ETHICS APPROVAL AND CONSENT TO PARTICIPATE This work is approved by the

Institutional Review Board of Jen-Ai Hospital in Dali, Taichung, Taiwan, under approval number 202200001B0. All participants signed an agreement form stating that the data obtained from the

experiment can be published in academic journals, with an ID that replaces their names. CHALLENGES OF COLLECTING SWING DATA Collecting data on table tennis swings presents a variety of

challenges. Commonly used methods, such as video recording or attaching sensors to the human body, have limitations. Video recording requires a fixed camera. It can be challenging to

replicate the same environmental setup from one place to another, making it difficult to collect accurate data in different environments. Attaching sensors to the body using smartwatches,

smartphones, or other sensors may influence players’ movements, creating interfering factors in the analysis. Embedding sensors into the equipment, such as the paddle, is a better option.

However, previous studies that used this approach collected only data from a few players, typically no more than 2012,13,14. In addition, a significant workforce is required to split the

continuous signals into stroke-based data. Given these limitations, we develop a new method that addresses the challenges of collecting swing data from table tennis. We embed sensors

directly into the racket to collect data from more than 90 professional players, generating a dataset of over 90,000 strokes. Our approach allows for accurately collecting stroke information

without significantly affecting players’ performance. Additionally, we have developed an automated method to split the collected continuous waveform data into each stroking data, making

data processing more efficient. By collecting data from many players, our method provides a more comprehensive understanding of the mechanics and nuances of table tennis swings. Our approach

overcomes the limitations of previous studies, which only collected data from a few players. The resulting dataset opens up new possibilities for research and development in the field,

allowing for the creation of more advanced applications and complex models. HARDWARE Figure 1 gives an overview of the entire system. Motion sensors are embedded in racket grips to collect

data. We use the shakehand grip style racket as it is more prevalent among players. The data collected are transferred to the RF wireless receiver, which transfers the information to a

computer through a USB port. Figure 2 shows how we embed the hardware in the table tennis racket. The embedded components include an inertial measurement unit (ICM-20948), a module for radio

frequency (RF) wireless transmission (E01-ML01SP), and affiliated components such as the button and the RGB LED for simple I/O communications, as shown in Fig. 3. A lithium battery powers

the system, and a TPS2546 USB charging port is connected to a 5V DC external power supply to maintain the battery’s power. Eventually, the embedded racket weighs approximately 190 grams,

which falls within the weight range of a regular racket. The motion sensor, ICM-20948, is critical to collecting swing data. The sensor integrates a 3-axis accelerometer, a 3-axis gyroscope,

and a 3-axis compass, forming a 9-axis sensor that can effectively measure 3-axis acceleration, 3-axis angular velocity, and 3-axis magnetic field data. The three axes are defined according

to Fig. 4: the positive x-axis is to the right, the positive y-axis is forward, and the positive z-axis is perpendicular to the red side of the racket. The motion sensors embedded in the

racket grips are configured with a sampling rate of 80 Hz, ensuring sufficient temporal resolution to capture detailed kinematic data. Calibration is performed using a standard vibration

exciter model 394C06 from PCB Piezotronics, ensuring the accuracy and reliability of the collected measurements. SWING DATA COLLECTION We invited 93 Taiwanese players from Group A to

participate in the data collection process. The players in Group A are elite players who have majored in physical education or have won medals in important competitions. Participants were

asked to perform at least one of three different swing modes using the proposed racket. The three modes include swing in the air, full power stroke, and stable hitting. Each mode requires

the participants to swing the racket 50 times continuously to generate a complete set of waveforms. In the full power stroke mode, the serving machine sets three ball speeds for players to

hit. SPLIT THE COMPLETE WAVEFORM SET INTO SEPARATED STROKE WAVEFORMS This section details the methodology for dividing a raw waveform into stroke-based waveforms. Before diving into the

specifics, we summarize the basic approach: the segmentation method involves integrating multi-axis signals into a single waveform, normalizing it to remove inconsistencies and noise, and

detecting peaks and troughs to isolate individual strokes. This approach, based on a combination of trend removal and peak detection, does not follow a prenamed standard but adopts

principles from signal processing tailored to this dataset. Table 1 shows the pseudocode of the split process. We describe the details in the following. As mentioned above, each participant

in the study swung the racket 50 times continuously, resulting in a complete set of waveforms. For further analysis, we want to divide each complete waveform set into 50 separate stroke

waveforms. However, this proved challenging, as different strokes exhibit different strengths and trajectories, generating unique waveforms. Figure 5a illustrates a portion of a complete

waveform set comprising ten consecutive strokes, each stroke waveform showing a similar but distinct shape. First, we integrate the six waveforms from the accelerometer and gyroscope into a

single _f_(_t_) by summing the absolute values of these waves, as shown by Equation (1). $$f(t)=| {A}_{X}(t)| +| {A}_{Y}(t)| +| {A}_{Z}(t)| +| {G}_{X}(t)| +| {G}_{Y}(t)| +| {G}_{Z}(t)| ,$$

(1) where _A__X_(_t_), _A__Y_(_t_), _A__Z_(_t_) are the 3-axis values from accelerometer at time _t_, _G__X_(_t_), _G__Y_(_t_), _G__Z_(_t_) are the values from gyroscope at _t_. We call the

outputted waveform the _integrated waveform_. Figure 5b displays an example of the integrated waveform. We normalize the integrated waveform as follows. First, we remove the trend from the

integrated waveform to remove the inconsistencies of each stroke from the same player. Next, we apply a low-pass filter provided by ICM-20948 to remove high-frequency noise. Finally, we

scale the waveform to be within the range of 0 to 1. These steps help to speed up the peak detection process in the subsequent steps. Figure 5c shows an example of the normalized waveform.

We segment the normalized waveform by stroke based on the following steps. We first plot a horizontal line _y_ = 1, which interacts with the peak of the entire normalized waveform. Since the

number of swings is known, we gradually move the horizontal line downward until the number of intersection points equals twice the number of swings. For example, in Figure 5d, the number of

known stroking features is 10, and the search is stopped when 20 intersection points are found. The peak of each stroke is expected to be within two neighboring intersections. Based on the

identified peaks, we further search for the nearest troughs to the left and right. These two troughs are split points for separating a complete stroke waveform. Figure 5e shows the two

troughs found for each wave, and Fig. 6 shows the 10 segmentation results. DATA RECORDS The dataset is available at Dryad15. This section introduces the released features and data

statistics. OVERVIEW OF THE RELEASED DATA The data released includes swing and personal features integrated into a tabular (CSV) format file. Table 2 shows the column headings. RELEASED

SWING FEATURES Table 3 details the three types of features derived from the waveforms. The first type includes the mean, variance, and root mean square of the accelerations and angular

velocities along the three axes (i.e., _A__X_(_t_), _A__Y_(_t_), _A__Z_(_t_), _G__X_(_t_), _G__Y_(_t_), _G__Z_(_t_)), which result in 18 features. The second type contains the mean, maximum

value, minimum value, skewness, and kurtosis of the overall acceleration _A_(_t_) and angular velocity _G_(_t_), which generates ten features. Finally, we apply the Fourier transform to the

acceleration and angular velocity waveforms and further derive the spectral density values and spectral entropy values for the acceleration and angular velocity, resulting in six features.

By converting the continuous waveform signals into a finite number of features, it should be more convenient to apply various machine learning and deep learning models for further analyses.

The unit of the accelerations (e.g., _A__X_(_t_), _A__Y_(_t_), _A__Z_(_t_)) is LSB/G (least significant bit per unit of G-force). By multiplying this value by 2/32768, the original G value

can be obtained. The unit of angular velocities (e.g., _G__X_(_t_), _G__Y_(_t_), and _G__Z_(_t_)) is LSB/deg/s (least significant bit per unit of angular velocity). By multiplying this value

by 250/32768, the original DPS (degree per second) can be obtained. RELEASED PERSONAL FEATURES We provide anonymized personal information for each player, including gender, age, height,

weight, handedness, racket-holding hand, and years of experience. These demographic details can be used for group comparisons, such as examining waveform characteristics across different

groups of players based on factors such as gender, dominant hand, or skill level. Additionally, to prevent attackers from recognizing a player’s identity from unique numerical features, we

categorized each numerical value into one of three labels – “low”, “medium”, or “high” – according to the feature’s distribution. DATA STATISTICS We recruited 93 players, comprising 53 males

and 40 females, and 78 are right-handed while 15 are left-handed. The statistical summary for other numerical features is listed in Table 4. Based on the swings of the 93 players, we

generate 97, 350 records: 7, 500 of them are mode 0 (swing in the air); 73, 850 of them are mode 1(full power stroke); 16, 000 of them are mode 2 (stable hitting). TECHNICAL VALIDATION The

TTSwing dataset was collected under controlled conditions to ensure the precision and reliability of the recorded data. The embedded 9-axis motion sensors were pre-calibrated, ensuring

precise accelerometer, gyroscope, and magnetometer readings. Furthermore, real-time wireless transmission of swing data to a laptop helped minimize possible data loss or corruption.

Demographic attributes such as gender, age, height, weight, and racket-holding hand were self-reported by participants, with verification when possible, such as verification from coaches or

retired athletes. Since these attributes are inherently factual and do not involve subjective measurement, they can be considered ground truth within the dataset. POTENTIAL SOURCES OF ERROR

IN DATA COLLECTION Several potential sources of error could affect the data collection process. First, variations in sensor calibration could introduce inconsistencies in the measurements.

Although embedded sensors were calibrated before data collection, slight drifts in sensitivity or accuracy may occur over time. Second, placement of the sensor in the racket grip, while

designed to minimize interference, can result in slight deviations due to minor changes during long-term use or repetitive impacts. Third, while efforts were made to automate waveform

segmentation, algorithmic errors during stroke identification may occasionally misclassify strokes or omit key features, particularly for players with unconventional playing styles. USAGE

NOTES We conduct experiments based on Python version 3.10. Packages and their tested versions are listed in Table 5. Once the code is downloaded from the repository, the users can use pip

install -r requirements.txt to reproduce the experimental environment. To run the code, users can change the directory to the src folder and run the Python scripts in the folder to reproduce

the results. CODE AVAILABILITY The code and data are available on Dryad at https://datadryad.org/stash/dataset/doi:10.5061/dryad.0zpc8677f15. REFERENCES * Iino, Y. & Kojima, T. Effect

of the racket mass and the rate of strokes on kinematics and kinetics in the table tennis topspin backhand. _Journal of Sports Sciences_ 34, 721–729 (2016). Article  PubMed  MATH  Google

Scholar  * Lees, A. Science and the major racket sports: a review. _Journal of sports sciences_ 21, 707–732 (2003). Article  PubMed  MATH  Google Scholar  * Iino, Y. & Kojima, T.

Kinematics of table tennis topspin forehands: effects of performance level and ball spin. _Journal of Sports Sciences_ 27, 1311–1321 (2009). Article  PubMed  MATH  Google Scholar  * Seong,

M. _et al_. Multisensebadminton: Wearable sensor–based biomechanical dataset for evaluation of badminton performance. _Scientific Data_ 11, 343 (2024). Article  PubMed  PubMed Central  MATH

  Google Scholar  * Wiles, T. M. _et al_. Nonan gaitprint: An imu gait database of healthy young adults. _Scientific Data_ 10, 867 (2023). Article  PubMed  PubMed Central  MATH  Google

Scholar  * Ozkaya, G. _et al_. Three-dimensional motion capture data during repetitive overarm throwing practice. _Scientific data_ 5, 1–6 (2018). Article  MATH  Google Scholar  * Ren, Y.,

Huang, Z., Guo, Y., Wu, J. & Sun, Y. Kinematic characteristics of backhand block in table tennis. In _Proceedings of the 4th International Conference on Biomedical Signal and Image

Processing_, 41–45 (2019). * Yanan, P., Jilong, Y. & Heng, Z. Using artificial intelligence to achieve auxiliary training of table tennis based on inertial perception data. _Sensors_ 21,

6685 (2021). Article  ADS  PubMed  PubMed Central  MATH  Google Scholar  * Viyanon, W., Kosasaeng, V., Chatchawal, S. & Komonpetch, A. Swingpong: analysis and suggestion based on motion

data from mobile sensors for table tennis strokes using decision tree. In _Proceedings of the 2016 International Conference on Intelligent Information Processing_, 1–6 (2016). * Sharma, M.,

Anand, A., Srivastava, R. & Kaligounder, L. Wearable audio and imu based shot detection in racquet sports. _arXiv preprint arXiv:1805.05456_ (2018). * Ferreira, N. M., Torres, J. M.,

Sobral, P., Moreira, R. S. & Soares, C. Classification of table tennis strokes in wearable device using deep learning. In _ICAART (3)_, 629–636 (2022). * Boyer, E., Bevilacqua, F., Phal,

F. & Hanneton, S. Low-cost motion sensing of table tennis players for real time feedback. _Int. J. Table Tennis Sci_ 8, 1–4 (2013). Google Scholar  * Blank, P., Hoßbach, J., Schuldhaus,

D. & Eskofier, B. M. Sensor-based stroke detection and stroke type classification in table tennis. In _Proceedings of the 2015 ACM International Symposium on Wearable Computers_, 93–100

(2015). * Blank, P., Groh, B. H. & Eskofier, B. M. Ball speed and spin estimation in table tennis using a racket-mounted inertial sensor. In _Proceedings of the 2017 ACM International

Symposium on Wearable Computers_, 2–9 (2017). * Chou, C.-Y. _et al_. TTSWING: a dataset for table tennis swing analysis, https://doi.org/10.5061/dryad.0zpc8677f (2024). Download references

ACKNOWLEDGEMENTS We acknowledge the support of the Taiwan National Science and Technology Council under grant number NSTC 112-2425-H-028-001. AUTHOR INFORMATION Author notes * These authors

contributed equally: Che-Yu Chou, Zheng-Hao Chen. AUTHORS AND AFFILIATIONS * Department of Computer Science and Information Engineering, National Central University, Taoyuan, 320317, Taiwan

Che-Yu Chou, Zheng-Hao Chen, Hung-Hsuan Chen & Min-Te Sun * Department of Computer Science and Information Engineering, National Formosa University, Yunlin, 63200, Taiwan Yung-Hoh Sheu *

Department of Sport Performance, National Taiwan University of Sport, Taichung, 404401, Taiwan Sheng K. Wu Authors * Che-Yu Chou View author publications You can also search for this author

inPubMed Google Scholar * Zheng-Hao Chen View author publications You can also search for this author inPubMed Google Scholar * Yung-Hoh Sheu View author publications You can also search

for this author inPubMed Google Scholar * Hung-Hsuan Chen View author publications You can also search for this author inPubMed Google Scholar * Min-Te Sun View author publications You can

also search for this author inPubMed Google Scholar * Sheng K. Wu View author publications You can also search for this author inPubMed Google Scholar CONTRIBUTIONS C.-Y.C., Z.-H.C. and

H.-H. C. wrote the paper. C.-Y.C., Z.-H.C. and H.-H.C conducted experiment(s) and analyzed the results. Y.-H.S. designed and implemented the hardware. Y.-H.S. and S.K.W. collected the data.

S.K.W. and M.-T.S. supervised the project. All authors reviewed the manuscript. CORRESPONDING AUTHOR Correspondence to Hung-Hsuan Chen. ETHICS DECLARATIONS COMPETING INTERESTS The

corresponding author is responsible for providing a competing interests statement on behalf of all authors of the paper. This statement must be included in the submitted article file.

ADDITIONAL INFORMATION PUBLISHER’S NOTE Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. RIGHTS AND PERMISSIONS OPEN

ACCESS This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format,

as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third

party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the

article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright

holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Reprints and permissions ABOUT THIS ARTICLE CITE THIS ARTICLE Chou, CY., Chen, ZH., Sheu, YH. _et

al._ TTSwing: a Dataset for Table Tennis Swing and Racket Kinematics Analysis. _Sci Data_ 12, 339 (2025). https://doi.org/10.1038/s41597-025-04680-y Download citation * Received: 05

September 2024 * Accepted: 19 February 2025 * Published: 26 February 2025 * DOI: https://doi.org/10.1038/s41597-025-04680-y SHARE THIS ARTICLE Anyone you share the following link with will

be able to read this content: Get shareable link Sorry, a shareable link is not currently available for this article. Copy to clipboard Provided by the Springer Nature SharedIt

content-sharing initiative