Background & Summary

Recent advances in wearable sensors have enabled the accurate recognition of human actions at a highly detailed level1,2. This, in conjunction with modern AI techniques, has facilitated the analysis of high-level actions in various fields, such as fall detection3,4,5,6,7, sports training8,9,10,11,12,13,14,15, healthcare16,17,18,19, assistive technologies for people with disabilities20,21, and rehabilitation22,23. While other fields focus on reducing the number and size of sensors used in research, the sports industry is adopting multimodal sensors to gain a comprehensive understanding of player movements and physical states. These sensors facilitate various analyses, from simple posture classification to performance analysis, and include inertial measurement units (IMUs)24,25,26,27,28,29, eye trackers30,51,52. These AI-assisted coaching methods offer a more accessible and inclusive approach to training, granting individuals 24/7 access to expert feedback49, or supplying an AI coaching service for those who might otherwise not have access to a coach36,38,40,45,47,53,54,55,56. Such ubiquity in training access ensures that a larger number of individuals, irrespective of geographical constraints or other limitations, can benefit from expert guidance and training programs.

Likewise, several AI- and sensor-based diagnostic systems have been proposed for badminton training8,9,10,11,12,13,14. Since badminton performance largely depends on the correct execution of each stroke, which requires quick and complex reflexes, sensor-based action analysis can be especially advantageous. Specifically, performing an effective badminton stroke requires proper stance, power control, and arm speed12, all of which are difficult for a human coach to monitor simultaneously. In particular, for beginners who are not yet familiar with basic badminton movements, acquiring the proper swing posture and power control can often require a considerable period of training, sometimes extending over several months. Therefore, by utilizing wearable sensors and AI technology to collect data from players of various skill levels, a system could be developed that not only assists in the training process but also provides an objective metric to complement a coach’s assessment.

Despite the benefits and prevalence of computer-assisted applications in badminton training, there is a limited amount of publicly accessible badminton action data available for training system development. Badminton datasets typically fall into two categories: individual stroke data collection in controlled settings, and strategy analysis based on real-world match videos (see Table 1). However, most publicly available datasets57,58,59,60 focus on match data between professional players, with an emphasis on tactical aspects. These aspects include predicting an opponent’s shuttlecock trajectory and stroke types59,60, as well as detecting strokes and identifying players’ bounding boxes57,58. The unpredictable trajectory, speed, and timing of the shuttlecock in real-match scenarios, coupled with the players’ dynamic movements, make modeling and evaluating individual strokes particularly challenging, as it requires accounting for previous strokes and the opponent’s actions.

Table 1 Comparison of the MultiSenseBadminton dataset[64] with existing public and non-public badminton datasets: In the “Context” column, “C” denotes collecting badminton data in a constrained, controlled environment, while “F” indicates data collection in the field during actual competitive play between two or more players.
Full size table

In contrast, individual stroke data collection in controlled setting, which is our primary area of focus, provides opportunities for stroke classification11,13,61,62, statistical comparisons across varying expertise levels9,63, and in-depth evaluation of each stroke14,62. Given the context in which the individual stroke data is used, gathering it in a stable environment is crucial, allowing to focus entirely on the mechanics of badminton movements and accurately capture the full dynamics. Specifically, the controlled environment for stroke data collection increases the potential to gather diverse biometric and motion data through wearable sensors. This includes IMU sensor data12,13,61,62,64, motion capture data9,63 as well as collecting biometric information like electromyography14,63,65 and foot pressure9. Such an environment also supports the use of cameras for analysis10,11.

However, there is a gap in the available datasets, particularly in terms of assessing the quality of badminton strokes. Existing research has not fully addressed key elements of badminton swings, such as the players’ skill levels, the final position of the shuttlecock after the swing, the quality of impact during the swing, and the hitting point. Furthermore, although most datasets focus mainly on data from one or two types of sensors, fully understanding player performance variations requires combining data from multiple sources, including motion, foot pressure, and muscle activity, to get a comprehensive view of stroke quality.

To address the gap identified in previous research, our study concentrated on evaluating the quality of individual strokes for players at various skill levels rather than focusing on tactical strategies. This approach involved collecting data in controlled environments, where players executed strokes using a shuttlecock launcher63 while being monitored with various sensors. Our research specifically concentrated on collecting data for two primary strokes taught in beginner badminton club courses: the forehand clear and the backhand drive14,66,67,68. We aimed to assess how the quality of each player’s stroke varies with different postures. For this, we collected over 150 swing data points per stroke type, involving players of various skill levels9. Therefore, we collected a multimodal badminton swing dataset incorporating both motion and physiological data, including full-body motion, foot pressure, gaze, and muscle activation data.

Our dataset surpasses previous approaches by incorporating a diverse range of data sources not included in existing badminton datasets. Our data collection setup, including the environment and sensor configuration, was developed based on insights from interviews with badminton experts. Our dataset encompasses five types of sensor data streams captured simultaneously, along with expert interviews, surveys, and annotated data. The dataset also includes video recordings from different point of view (Front, Side, Whole, Eye, and Eye with Gaze Overlay). The annotation data includes stroke type, skill level, ball landing location, shuttlecock sound, and hitting position. By compiling this comprehensive dataset, we provide a detailed representation of badminton strokes and related characteristics. This dataset can be leveraged to develop training programs, performance analysis techniques, and coaching strategies in the sport of badminton.

Our research additionally introduces an initial framework for utilizing machine learning with our dataset in the Technical Validation section. This section outlines a methodology that includes preprocessing and feature extraction, emphasizing the suitability of our dataset for machine learning applications. This includes providing examples of classifying stroke type, skill level, horizontal and vertical landing position, hitting point, and stroke sound. We reported the accuracy of our annotations with state-of-the-art machine-learning techniques. To facilitate usage by a wide audience, including those not specialized in deep learning, we have provided examples and made our deep learning pipeline source code openly accessible on the project’s GitHub page.

Methods

Dataset design

To build a badminton action dataset designed specifically to address the needs of the badminton coaching field, we engaged three professional badminton coaches from a local club that boasts a membership of over 50 individuals. Each coach had undergone professional training and had a minimum of five years of coaching experience (Female: 1; Age: Mean = 36.7, SD = 11.3; Years of Experience: Mean = 11, SD = 5.1). Our main aim was to extract their knowledge, focusing particularly on insights related to the overall training process. This knowledge would then guide us in selecting an appropriate sensor set and designing a dataset to facilitate player performance analysis and feedback.

We centered our discussions around critical elements that expert coaches pay attention to when teaching swing techniques. This included their standardized training processes, strategies for providing feedback, and the strategies used for executing stroke actions. To prepare for the interviews, we sent the questions to the coaches in advance. Each interview lasted roughly an hour, and each coach received a compensation of $80 for their participation.

The following subsections provide an aggregated summary of the interview responses. While our dataset entails the complete answers to a total of six questions, this paper specifically highlights the four questions that directly influenced the design of the dataset (see Table 2). These four questions primarily focus on the types of data and annotations that should be included for the analysis of badminton stroke actions. The full set of six questions, providing comprehensive insights into the coaches’ views on the applicability of AI-based coaching systems, challenges faced in current training methodologies, and thoughts on coaching within a virtual environment, can serve as a useful guide for future dataset design efforts.

Table 2 Summary of Interview with Badminton Coaches.
Full size table

Question 1. What is the most important skill to teach during badminton training?

Summary of responses to Question 1

The coach highlights the significance of grip, posture, swing, and step in badminton training. Grip training is emphasized as a continuous process to enhance racket control and precision. For beginners, the coach prioritizes teaching proper posture, followed by improving swing accuracy and shot execution. Maintaining the correct swing involves generating power from the rotation of the torso, along with the coordinated movement of the arm and wrist. The evaluation of shot accuracy focuses on hitting the intended target or clearing the net correctly. In terms of step, the coach emphasizes positioning the dominant foot underneath the shuttlecock’s expected landing spot and highlights the importance of the split step for quick post-shot preparation. Overall, the coach underscores the importance of these elements in effective badminton training.

Question 2. What is the criterion for evaluating the success of badminton training?

Summary of responses to Question 2

According to the interviews, coaches evaluated the badminton training process based on several key factors. First, the point of impact at which the ball hits the racket is evaluated-specifically, whether the ball is in front of the body when hitting. Second, the trajectory and landing location of the ball are assessed to ensure that it travels through the target distance and direction. Third, the accuracy of the stroke, which is determined by assessing whether the ball makes contact with the center of the racket and the sound produced during this interaction, is evaluated. Finally, the speed of the ball is also monitored.

Question 3. How do you give feedback to trainees during training?

Summary of responses to Question 3

Coaches typically employ four main methods to provide feedback to badminton students. First, the coaches provide verbal feedback to the students to inform them of their performance and whether they have executed the correct stroke. If a student still has difficulty maintaining the proper posture or executing the correct movement, the coach may demonstrate an example to clarify the appropriate form. Additionally, some coaches may use video feedback to help students track their progress or observe the correct motion of a skilled player. However, while video feedback can be effective for some students, it may not always be the most useful tool for everyone. Although videos can provide a visual aid for learning, many students may find it difficult to fully grasp the technique without the opportunity to physically practice and experience the movements themselves. In particular, providing feedback on concepts that are difficult to understand visually, such as the application of force or shifting one’s center of gravity, can present challenges for coaches. Therefore, coaches may need to tailor their feedback strategies to suit individual learning styles and preferences to optimize student learning and development during badminton training.

Question 4. What are the important data for an effective badminton stroke?

Summary of responses to Question 4

Several factors contribute to the effectiveness of a badminton shot, including swing accuracy, footwork, and gaze processing, all of which collectively help to ensure the shuttlecock is hit at the optimal timing and trajectory. Executing a fast and precise swing entails a sequence of actions: visually tracking the ball, extending the arms, step** towards the target, and delivering a forceful impact. Holding the racket correctly is also crucial in badminton to generate impact during a shot, enabling greater wrist flexibility and range of motion to execute various types of shots. Proper footwork is essential for maintaining appropriate body positioning and stable shots, relying on precise balance control and efficient movement patterns, which should be attentively practiced to direct the shuttlecock towards the intended direction at the desired speed. Consequently, sensor-based analysis of badminton strokes requires the collection of data on various factors, including tracking racket position, monitoring ball trajectory, measuring hand pressure, tracking eye movements, monitoring body and foot positioning, assessing foot pressure, and analyzing muscle activity.

Interview-based dataset design

In our study, we established the sensor set, data collection environment, target strokes, and annotation data through interviews with experts. The selection of the most suitable sensors, guided by insights from Question 4 in the interviews, emphasized the need for sensors that capture crucial data without impeding natural badminton swing movements. Therefore, we opted for a non-invasive, comprehensive sensor set suitable for players of various skill levels, including eye gaze tracking, EMG, IMU-based body tracking, and foot pressure sensors.

To avoid altering the racket’s weight or feel with attached sensors, we chose motion tracking technology worn on the hand (Perception Neuron studio) to measure racket movements. This decision was informed by studies in racket sports69,70,71,72,73, where IMU sensors on the hand provided stroke classification performance comparable to sensors on the racket74, and in some cases, even superior correlation with player performance-related measures72. This approach allows us to gather essential swing information via IMU sensor-based data from the hand, maintaining the racket’s natural feel and balance during play and offering proxy measures for racket dynamics.

For the data collection setting, we drew inspiration from typical badminton training environment where a coach throws shuttlecocks for the trainee to return, often providing real-time posture correction and feedback. To replicate this training environment consistently in our study, we employed a shuttlecock launcher for collecting badminton stroke data63. We calibrated the launcher to consistently release shuttlecocks at the same angle for each stroke type, which enabled the collection of comparable swing data across different participants. This setup was instrumental in gathering data on how players of various skill levels respond to the same shuttlecock trajectory, thereby facilitating an analysis of the diversity in players’ responses to uniform strokes. To thoroughly observe the participants’ posture and the shuttlecock’s trajectory, we installed three external cameras, each capturing a unique view. In addition, an eye tracker camera was utilized to record sound.

For our target strokes, we concentrated on two basic strokes essential for beginners in badminton training: the forehand clear and backhand drive. This focus was driven by the goal of our dataset, which is to evaluate the quality of individual strokes for building a badminton training system. By narrowing down to these fundamental strokes, we aimed to collect over 150 data points per participant, focusing on how players of different skill levels react to shuttlecocks with the same trajectory and how their responses vary. Drawing on answers from Question 2 of our expert interviews, we established criteria to evaluate each badminton stroke and annotated these criteria to build a dataset on stroke quality. Our annotations included aspects such as the skill level of each player, the horizontal and vertical location of each strokes, the hitting point, and the sound quality produced during the hit. By concentrating on these specific strokes and detailed annotations, we sought to provide a comprehensive dataset that would offer insights into the nuances of stroke execution and quality across various skill levels in badminton.

Ethics statement for the multisensebadminton dataset

The development of the MultiSenseBadminton dataset received ethical clearance from the Institutional Review Board (IRB) at the Gwangju Institute of Science and Technology75. This project was approved under the protocol code 20220628-HR-67-20-04 on July 21, 2022.

Upon arrival at the data collection site, participants were presented with consent forms. These forms required thorough reading and written agreement from participants, confirming their willingness to contribute to the data collection process. A critical aspect of the MultiSenseBadminton dataset is its public availability. As such, explicit consent was also obtained for the public release of data that includes personally identifiable information (PII), specifically the video recordings of participants performing badminton swing. The videos have undergone mosaic or blur processing for participant confidentiality. The final MultiSenseBadminton dataset comprises video, annotation, personal information and sensor data from all 25 participants who participated in the study, each of whom consented to the release of their personal data for public access.

Participants

In this study, data were collected from 25 participants (20 males and 5 females) aged between 18 and 52 years (Mean = 26.8 years, SD = 6.59 years). The basic physical conditions of the participants were as follows: weight 48–108 kg (Mean = 76.4 kg, SD = 14.6 kg); height 160–190 cm (Mean = 174 cm, SD = 8.33 cm) (Shown in Table 3). The participants’ training experience varied between 0 and 22 years (Mean = 3.96 years, SD = 6.9 years). All participants demonstrated dominance in one hand and confirmed that they predominantly use this hand in badminton. Before data collection, the participants were briefed about the use of multiple wearable sensors and agreed to participate in the study. After collecting data, the participants were paid $40 for participating. All participants consented to data disclosure, and data for all subjects were included in our dataset.

Table 3 Demographics and Training Experiences of Subjects: The self-reported skill level was recorded on a 7 Likert scale; the higher the Likert value, the higher the skill level.
Full size table

Sensors and data collection framework

The sensor collection framework used in our study is an adaptation of the ActionSense framework76. The original framework encompassing codes, a graphical user interface (GUI), and sensor visualization capabilities was tailored for human-activity data collection from wearable devices during kitchen activities. We therefore modified this framework to align with our unique sensor set and dataset design. This involved customizing the GUI and real-time data visualization features of ActionSense for in-situ monitoring and time-synchronized annotation during data collection.

Our study utilized five types of wearable sensors: eye tracking, body tracking, foot pressure, and EMG sensors. We supplemented these with three cameras and a shuttlecock launcher for comprehensive data collection focused on badminton stroke analysis (see Fig. 1). Each sensor’s data stream was integrated into the overarching ActionSense framework by connecting their respective Python API or stream layer API, thereby facilitating data import via TCP/IP communication.

Fig. 1
figure 1

The sensors used for data collection during the experiment.

Full size image

To ensure easy data manipulation and seamless integration, we opted to save the collected data in an HDF5 file format77. This format offers cross-platform and cross-language compatibility, rendering it a versatile choice for storing and accessing large volumes of scientific data. Furthermore, the HDF5 format allows for the hierarchical organization of multimodal heterogeneous data such as sensor readings. Given its capability for processing large data in concurrent threads and parallel I/O, the HDF5 format is particularly suitable for our data-rich configurations that involve five wearable sensors and three cameras in the simultaneous data stream channels.

We collected all data using Unix time to assist in analyzing the temporal relationship between different sensor data points. The structure of the sensor acquisition framework is illustrated in Fig. 2. Notably, even in the event of a sensor disconnection during data collection, the entire framework continues data acquisition seamlessly.

Fig. 2
figure 2

Sensor Data Collection Framework.

Full size image

Eye tracking (Pupil Invisible Glasses)

In our study, we utilized Pupil Invisible glasses to collect 1) information on whether gaze data were successfully received; 2) 2D gaze data (horizontal and vertical); 3) ambient sound that participants were subject to; and 4) first-person videos through eye camera. The Pupil Invisible glasses are a wearable eye-tracking system designed to resemble a regular pair of glasses. The system includes two inner cameras on the frame to track eye movements, an exterior camera on the left temple with a wide field of view to record the environment, and a USB-C connector on the right temple that connects to a smartphone running the tracker application. The eye-tracker output comprises recorded videos displaying the participant’s gaze position and coordinates relative to the outside image. While the Pupil Invisible glasses autonomously estimate gaze data via an embedded convolutional neural network algorithm6. This approach tests the model’s ability to generalize to new subjects across different skill levels. Additionally, for comparison, we developed a baseline model that predicts the predominant class, serving as a benchmark to assess our models’ performance and effectiveness against this basic approach.

Table 6 Leave-Three-Out reference subject number; This table provides the reference subject numbers for the LTO cross-validation process, facilitating a reproducible benchmark.
Full size table

Stroke type and skill level classification results

In the case of stroke type classification, the study involved categorizing three types of strokes (forehand clear, backhand drive, and non-stroke). Table 7 displayed the mean and standard deviation of accuracy, balanced accuracy, and F1 score, obtained through LTO validation. Overall, deep learning models outperformed the baseline in all metrics, with ConvLSTM demonstrating particularly superior performance across all metrics compared to other models.

Table 7 Stroke Type Classification Results; models with the highest performance in each metric are highlighted in bold.
Full size table

In the case of skill level classification, the study involved categorizing three types of strokes (Beginner, Intermediate, and Expert), shown in Table 7. Overall, deep learning models outperformed the baseline in all metrics, with LSTM demonstrating particularly superior performance across all metrics compared to other models.

Annotation classification results for clear

In the classification of annotation data related to clear strokes, Table 8 presents an analysis results using different models, evaluated through both LTO and 10-fold cross-validation methods. In the horizontal landing location classification, the Transformer model stood out in the LTO results, with its efficacy closely matched in the 10-fold setting. In the vertical landing position classification, the baseline model unexpectedly outperformed others in the LTO approach. This could be attributed to the vertical landing position being relatively consistent across players of different skill levels, leading to higher baseline accuracy as the data for this category is less variable. Additionally, in the hitting point classification, a similar trend was observed where the baseline accuracy was significantly high, likely due to the predominant distribution of data being classified as “Front”, skewing the results. However, in the 10-fold cross-validation for vertical landing position, the LSTM model demonstrated superior performance. The hitting point classification showed strong performances from both the ConvLSTM and LSTM models, with the latter slightly edging out in the 10-fold cross-validation. Lastly, in the sound classification, the ConvLSTM model exhibited the best performance in all metrics in the LTO approach, while in the 10-fold cross-validation, the LSTM model led in both accuracy and F1 score.

Table 8 Clear Classification Results; models with the highest performance in each metric are highlighted in bold.
Full size table

Annotation classification results for drive

Table 9 presents results across different models, evaluated using both LTO and 10-fold cross-validation. In the horizontal landing position classification, the performance varied between the LTO and 10-fold settings. In the 10-fold cross-validation, the Transformer model achieved the highest accuracy, while the ConvLSTM model had the highest balanced accuracy, and the LSTM model led in F1 score. In contrast, in the LTO results, the baseline model achieved the highest accuracy, ConvLSTM led in balanced accuracy, and the Transformer model scored the highest in F1 score. For the vertical landing position classification, the baseline model’s superior performance in the LTO approach can be attributed to the relatively consistent vertical landing positions across different players, resulting in less variable data and therefore higher baseline accuracy. In contrast, in the 10-fold cross-validation, the LSTM model showed the highest balanced accuracy and F1 score for vertical landing position. In the hitting point classification, the majority of the data was distributed in the “Front” category. This predominance led to no significant performance difference between the baseline and the top-performing models across two validation settings. Lastly, in the sound classification, varied performances were observed among the models. The Transformer model excelled in the 10-fold cross-validation, while the ConvLSTM and LSTM models displayed strong results in the LTO settings, particularly with LSTM leading in LTO for F1 score and accuracy.

Table 9 Drive Classification Results; models with the highest performance in each metric are highlighted in bold.
Full size table

The provided pipeline represents a preliminary implementation of a deep-learning application to evaluate the suitability of the MultiSenseBadminton sensor and annotation dataset75. An important consideration in our study is the imbalance in data distribution across participants, particularly evident in the annotations for horizontal landing position, vertical landing position, and hitting point. This imbalance inherently leads to higher baseline accuracy in certain cases, as the majority class tends to dominate the dataset. This phenomenon is especially notable in the vertical position and hitting point classifications, where the baseline accuracy occasionally surpasses that of the deep learning models. Such an outcome underscores the challenges posed by imbalanced data in machine learning, especially in the context of sports analytics where participant variability can significantly impact model performance. It highlights the need for careful consideration of data distribution and participant variability when interpreting model accuracy and effectiveness in classifying stroke-related annotations.

Usage Notes

Sensor-data visualization

We developed a data visualization tool to visualize sensor data and enable visual comparison of strokes, as shown in Fig. 13. The tool code is available on the project’s GitHub page, providing the capability to select various parameters for comparison, such as participant number, stroke type, stroke number, and sensor. These features allow for a comprehensive analysis of stroke patterns by facilitating the comparison of sensor data across different participants and strokes.

Fig. 13
figure 13

Data Visualizer.

Full size image

Data limitations

The MultiSenseBadminton dataset75 has certain limitations that need to be addressed. First, some of the wearable sensors used in this study were susceptible to noise, drift, and connectivity issues during data collection. For instance, the body tracking sensor caused position drift, which is a common problem with IMU-based motion tracking sensors90,91. Although we mitigated this issue by calibrating the motion tracking sensors for each session and recalibrating them upon detection of discrepancies between participants’ actual movements and the corresponding joint data, the inherent nature of the IMU sensor inevitably introduces some drift-related error. Further, the Cognionics AIM sensor generated spike values during the stroke, and we therefore wrote preprocessing code to process the spike values. The preprocessing code has been uploaded to the GitHub page. There was also an interruption in the streaming of data from the gForce EMG armband during the data collection process. To address this issue, the badminton stroke was stopped and the stream was reconnected to resume data collection. This ensured that the collected data corresponded to actual badminton strokes and were free from interruptions or inaccuracies.

The second limitation of the MultiSenseBadminton dataset75 is that the data were collected in a constrained environment, designed to mimic a typical badminton coaching scenario. In this setup, participants responded to shuttlecocks launched by a machine, similar to how a coach might throw shuttlecocks to a trainee. This method ensured that shuttlecocks were delivered in a uniform trajectory, allowing us to gather consistent sensor data across various aspects such as movement, muscle activity, and center of pressure shifts among players of different skill levels. While this approach is useful for controlled training exercises, future studies should aim to collect data in real-world match environments. In this real-match scenario, it’s essential to use wearable sensors that do not restrict the participants’ movement. Therefore, we recommend utilizing non-intrusive sensors, such as insole-based foot pressure sensors and cameras, to ensure free and natural player movement. Additionally, for a more effective and realistic data collection, it is advisable to recruit participants with intermediate or higher skill levels who are capable of engaging in actual gameplay. This method of data collection in real-match settings will provide insights into player strategies and movements, enhancing the understanding of competitive badminton dynamics.

Third, the location of the wearable sensor presented limitations that prevented the collection of whole-body data on badminton strokes. The sensor was attached to a specific part of the body, which restricted data collection to that particular area. As a result, we were unable to obtain a comprehensive understanding of the mechanics of badminton strokes across the entire body. For instance, the use of wearable sensors in this study was limited by the number of available sensors, and as a result, EMG sensors were only attached to the dominant arm and foot. Moreover, the discomfort associated with wearing multiple sensors can affect the accuracy of badminton stroke data. This is a significant limitation of wearable sensors that needs to be addressed. In fact, some participants in the study reported various discomforts when wearing eye-tracking glasses and EMG armbands. They also reported slightly different experiences when sporting sensors during badminton motion than when not wearing any. Therefore, after collecting such data, it is essential to conduct additional research to reduce the number of sensors by analyzing the correlations between them.

Fourth, the use of video-based annotation limits the accuracy of annotation data. We used three cameras to record participants from different angles during data collection. As it was difficult to record aspects such as the hitting sound, hitting point, and landing position during the data collection process, the annotations for Levels 3, 4, and 5 were conducted after the data collection by three annotators. The inter-rater reliability was assessed to ensure the accuracy of the annotations. The inter-rater reliability value for Level 3 was relatively low owing to the subjective nature of the hitting sound. Moreover, even in the case of Levels 4 and 5, some annotations had missing values, or the agreement between annotators was low. To overcome these limitations, future research should focus on develo** a system that can automatically detect ball trajectories and hit points. This would significantly improve the accuracy and reliability of annotation data and provide researchers with more comprehensive insights into the mechanics of badminton strokes.

Finally, our dataset focused on only two strokes, the forehand clear and backhand drive. Since our dataset’s objective was to assess the quality of badminton strokes across all skill levels, from beginners to experts, we concentrated on these two fundamental strokes. However, we recognize that focusing exclusively on these two strokes is a limitation. Moving forward, it is crucial to target players at intermediate levels and above, who are proficient in a broader range of advanced badminton techniques. Our future aim is to build a dataset for evaluating the quality of strokes involving more complex techniques such as hairpin, net shots, smashes, and drop shots, thereby expanding the scope and utility of our research in badminton stroke quality assessment.