BadminSense: Enabling Fine-Grained Badminton Stroke Evaluation on a Single Smartwatch
Authors: Taizhou Chen, Kai Chen, Xingyu Liu, Pingchuan Ke, Zhida Sun
Organizations: Department of Computer Science, Shantou University, Shantou, China · CSSE, Shenzhen University, Shenzhen, China · Department of Sociology, Hong Kong Shue Yan University, Hong Kong, China
Evaluating badminton performance often requires expert coaching, which is rarely accessible for amateur players. We present BadminSense, a smartwatch-based system for fine-grained badminton performance analysis using wearable sensing. Through interviews with experienced badminton players, we identified four system design requirements with three implementation insights that guide the development of BadminSense. We then collected a badminton strokes dataset on 12 experienced badminton amateurs and annotated it with fine-grained labels, including stroke type, expert-assessed stroke rating, and shuttle impact location. Built on this dataset, BadminSense segments and classifies strokes, predicts stroke quality, and estimates shuttle impact location using vibration signal from an off-the-shelf smartwatch. Our evaluations show that BadminSense achieves a stroke classification accuracy of 91.43%, an average quality rating error of 0.438, and an average impact location estimation error of 12.9%. A real-world usability study further demonstrates BadminSense's potential to provide reliable and meaningful support for daily badminton practice.
Figures & tables
Figure 1. Illustration of the BadminSense’s concept. The system leverages a single smartwatch to detect and provide fine-grained badminton stroke analysis, including stroke classification, stroke quality rating, and impact location estimation. Three screenshots on the right illustrate the front-end interface of BadminSense. Left: Session overview. This interface presents the stroke count in a game session with a performance summary. Middle: Session timeline view ( DR4 ). This interface shows the sequence of strokes executed throughout the session period ( DR4 ). Right: Stroke view. This interface displays the stroke quality rating ( DR1 ), the impact location ( DR2 ), and the improvement advice ( DR3 ) from top to bottom.
Figure 2. Stroke-specific impact location that most users prefer. The light green area indicate s the optimal impact location.
Figure 3. (a) Configuration of the badminton court for data collection and the data collection procedure. 1. Facilitator 1 serves a shuttle sprayed with ink, 2. Subject returns the racket after stroking, 3. Facilitator delivers a cleaned racket for the next stroke, 4. Facilitator 1 records the ink mark and cleans up the racket, 5. Facilitator 2 prepares a clean racket for the next stroke. The red dot in the top left corner denoted the smartphone camera recording position for ground truth labeling. (b)(c) Shuttle with fountain pen ink on the head before serving. (d) An ink mark on the racket face by the shuttle impact after stroking, highlighted in a red circle.
Figure 4. Illustration of the four stroke types that we focused on in this study. The corresponding bar chart below visualizes the distribution of the number of ratings, with the digit below indicating the total sample size in the category.
Figure 5. BadminSense system pipeline overview. The pipeline consists of four components: stroke segmentation, stroke classification, stroke quality rating, and impact location estimation.
Figure 6. (a) Illustration of the smartwatch coordinates. (b) Gyro data sample of each stroke from one user.
Figure 7. Pair-wise illustration of five sequential phases on a forehand overhead clear stroke sample. The five-phase segments in (a) are produced with a corresponding phase motion as illustrated in (b).
Figure 8. The confusion matrix of the SVM results. BOC, FOC, FOS, and FOD denote backhand overhead clear , forehand overhead clear , forehand overhead smash , and forehand overhead drop , respectively.
General
User-Independent
Accuracy
Percision
Recall
F1 score
Accuracy
Percision
Recall
F1 score
SVM
95.05%
95.39%
95.28%
95.23%
91.43%
92.03%
91.99%
92.00%
Linear Regression
54.95%
47.84%
39.69%
41.88%
41.90%
40.53%
28.66%
31.74%
Random Forest
79.72%
81.19%
80.47%
80.29%
66.19%
75.43%
67.62%
66.63%
MLP
92.57%
93.03%
92.91%
92.88%
80.95%
82.47%
81.86%
82.00%
Table 1. Stroke classification performance across models and training strategies
General
User-Independent
BOC
FOC
FOS
FOD
AVG
BOC
FOC
FOS
FOD
AVG
SVR
0.420
0.404
0.413
0.456
0.423
0.411
0.426
0.432
0.483
0.438
LinearReg
1.269
0.542
0.576
0.605
0.748
0.757
0.771
0.553
0.703
0.696
CNN-LSTM
0.574
0.534
0.651
0.513
0.568
0.700
0.727
0.595
0.474
0.624
MS-TCN
0.773
0.574
0.661
0.668
0.669
0.578
0.484
0.516
0.683
0.565
Transformer
0.658
0.540
0.601
0.560
0.590
0.802
0.571
0.669
0.763
0.701
Table 2. MAE of stroke rating performance across models and training strategies. A lower MAE indicates better results. BOC, FOC, FOS, and FOD denote backhand overhead clear , forehand overhead clear , forehand overhead smash , and forehand overhead drop , respectively.
Figure 9. The multi-model signal processing pipeline for evaluating impact location estimation performance. We evaluate this pipeline by testing across multiple machine learning regressors and deep learning encoder models. The processing flow with a thick green line yields the base performance on our task, with an SVR model.
Figure 10. Visualization of the impact location estimation results.
Feature
General
User-Independent
X
Y
AVG
X
Y
AVG
SVR
IMU
0.136
0.129
0.132
0.133
0.124
0.129
Acoustic+IMU
0.135
0.130
0.132
0.135
0.133
0.134
LinearReg
IMU
0.346
0.320
0.333
0.336
0.263
0.300
Acoustic+IMU
0.168
0.159
0.164
0.162
0.173
0.168
CNN
IMU
0.136
0.134
0.135
0.134
0.124
0.129
Table 3. The normalized MAE of the impact location estimate performance on X and Y axis across models, feature selection strategies, and training strategies. A lower MAE indicates better results.
Figure 11. The detailed questionnaire results distribution from the user evaluation. The digit on the bar indicates the number of ratings.
Multimodal resources for non-periodic court sports with laboratory-grade sensing remain scarce: few publicly pair instrumented ground reaction force (GRF) with high-frame-rate multi-view video, limiting markerless load estimation in realistic training settings. BadmintonGRF records eight synchronized RGB views at ~120 FPS, four Kistler force plates, and Vicon motion capture (C3D) without hardware genlock across modalities; alignment combines human-verified events, automated quality assurance, and per-camera time offsets with uncertainty metadata. Tier 1 distributes pose, time-aligned GRF, metadata, and splits under CC BY-NC 4.0, enabling the primary benchmark without raw RGB or C3D; we report a Tier 1 task that maps 2D pose to GRF. Tier 2 provides raw RGB and C3D under controlled access for studies that require appearance or full kinematics. The public release contains 17,425 impact-segment archives in the 10-subject benchmark tree (156 instrumented trials; raw multi-view RGB alone exceeds 1 TB); benchmark loader gates retain 12,867 view-specific instances and 1,732 unique impacts after multi-view deduplication. We are not aware of prior public badminton corpora that combine this sensing layout with audited video--GRF alignment for impact-centric GRF estimation. We distribute preprocessing code, leave-one-subject-out splits, ten reference baselines, and optional late fusion (one deterministic test-time pass per instance; no test-time augmentation), with a within-trial diagnostic in the supplementary material.
Kuoye Niu, Jianwei Li, Shengze Cai +6
Beijing Sport University · Beijing, China · Zhejiang University +4
Quantifying impact phenomena in badminton smashes is important for evaluating both athletic performance and equipment; however, conventional measurement systems involve trade-offs between temporal resolution, data efficiency, and preparation effort. This study proposes a measurement method using two synchronized event cameras to automatically estimate impact time, impact location on the racket face, and post-impact shuttlecock speed in an integrated manner within the same trial. The swing interval was detected from event rate statistics, impact time was estimated from the shuttlecock trajectory inflection in the lateral-view event data, impact location was determined by ellipse fitting to the racket face in the rear-view event image, and shuttlecock speed was calculated in the sagittal plane. To validate the proposed method, Bland-Altman analysis was performed against a high-speed camera-based reference method using 125 smash trials from five players. Impact time and shuttlecock speed were estimated in all 124 analyzable trials, and impact location was estimated in 93.5% (116/124). The bias (95% CI) for impact time, medio-lateral impact location, longitudinal impact location, and shuttlecock speed were 1.84 ms (1.45 to 2.23), 3.45 mm (2.18 to 4.72), -1.92 mm (-2.97 to -0.88), and -1.00 m/s (-2.46 to 0.46), respectively. No proportional bias was observed for any metric. These results suggest that the proposed method can serve as a useful tool for integrated assessment of badminton smash performance and equipment in practical settings.
Yudai Washida, Yuto Kase, Kai Ishibe +2
MIZUNO Corporation, 1-12-35 Nanko-Kita, Suminoe-ku, Osaka-shi, Osaka 559-8510, Japan
Accurate badminton stroke prediction is crucial for fine-grained sports analysis and tactical decision support. However, existing methods struggle to model rich temporal context. This paper introduces TemPose-TF-ASF (Adjacent-Stroke Fusion), a context-aware extension of TemPose. It enhances stroke recognition by incorporating stroke-type information from both preceding and subsequent strokes. A two-stage training and inference strategy is adopted. Preliminary predictions from the baseline model are reused as estimated temporal context. These predictions guide the joint optimization of the ASF module and the classifier. By explicitly modeling bidirectional temporal stroke dependencies, the proposed method can be seamlessly integrated into existing state-of-the-art models. Experiments on a large-scale badminton match dataset show consistent improvements over the baseline and its variants in terms of Accuracy and Macro-F1. Moreover, integrating ASF into other advanced methods yields notable performance gains. These results demonstrate strong transferability and generalization capability.
Tzu-Yu Liu, Duan-Shin Lee
Institute of Communications Engineering National Tsing Hua University, Hsinchu City, Taiwan · Department of Computer Science National Tsing Hua University, Hsinchu City, Taiwan