Badmintonvision: a deep learning framework for automated tactical analysis in elite badminton

Abstract

Tactical analysis in badminton is critical for performance optimization, yet traditional manual video review is time-consuming, subjective, and difficult to scale. Existing deep learning approaches struggle with rapid stroke recognition from single frames and require extensive labelled data. We present BadmintonVision, an integrated framework combining self-supervised learning (SSL), temporal modelling, and tactical analysis for automated badminton analysis. The detection model is pretrained on 139,501 unlabelled images using ConvNeXt V2. We develop MotionFormer, a Transformer-based module processing 10-frame windows (400 ms) to capture complete stroke cycles and construct a dataset of 30,000 expert-annotated images from BWF World Championships (2019–2025) covering five action classes. BadmintonVision achieves 94.1% detection accuracy (mAP@0.5), with SSL pretraining providing a 9.8% relative gain in stroke recognition and temporal modelling providing an additional 6.3% relative gain. Movement analysis reveals that scoring rallies correlate with higher path efficiency (Cohen’s d = 1.77) and shorter recovery times (d = − 1.56), though these correlational findings warrant cautious interpretation. In an exploratory, gender-stratified analysis, Markov-chain modelling further suggests preliminary, gender-associated regularities in stroke-transition sequences; given the limited number of games and players analysed, these patterns are descriptive and warrant confirmation in larger, balanced samples. BadmintonVision offers a reproducible approach for quantitative tactical analysis supporting coaching decisions and performance research in elite badminton.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13102-026-02073-1.

Introduction

Badminton is one of the fastest racket sports, with shuttlecock speeds exceeding 400 km/h during elite competition [1]. Its combination of extreme speed and tactical complexity has made badminton an active subject of sports-science and performance-analysis research [1, 2]. Performance optimization in this sport requires detailed understanding of player movements and stroke patterns across different match situations. Traditional analysis relies on manual video review by coaches and analysts who watch match recordings to identify technical errors and tactical patterns. This approach has three main limitations: it is time-consuming because a single match may require several hours of review; it is subjective because different analysts may interpret the same actions differently; and it has limited scalability because the workload increases linearly with the number of matches and players to analyse [3, 4].

Recent advances in computer vision and deep learning have enabled automated sports video analysis. Object detection methods such as YOLO (You Only Look Once) have been applied to detect players and track movements in various sports [5, 6]. TrackNet was developed specifically for tracking small and fast-moving objects like shuttlecocks in broadcast videos [7, 8]. Several studies have explored badminton action recognition using convolutional neural networks and pose estimation techniques [9–11]. However, current approaches face three challenges. First, single-frame detection methods struggle to recognize rapid stroke actions that span only 2–4 frames at typical broadcast frame rates [12]. Second, most models are trained with limited labelled data because expert annotation of sports actions is expensive and time-consuming [13]. Third, existing systems focus on individual components such as player detection or shuttlecock tracking but lack integration for comprehensive tactical analysis [14, 15].

To address these challenges, we present BadmintonVision, a deep learning framework for automated tactical analysis in elite badminton. Our framework has three main components. First, we apply self-supervised learning (SSL) using ConvNeXt V2 [16] to pretrain the detection model on 139,501 unlabelled badminton images, which improves stroke recognition accuracy by 9.8% compared to supervised-only training. Second, we develop MotionFormer, a Transformer-based temporal module that processes 10-frame windows (400 ms) to capture complete stroke cycles including preparation, execution, and recovery phases. Third, we integrate downstream analysis pipelines including player trajectory extraction, movement speed calculation, and Markov chain modelling of stroke sequences. Applied to broadcast footage from BWF World Championships (2019–2025), our framework achieves 94.1% detection accuracy (mAP@0.5) for stroke recognition and reveals quantitative differences in movement patterns and tactical signatures between scoring and conceding rallies. BadmintonVision provides an efficient and reproducible approach for tactical analysis that can support coaching decisions and performance research in elite badminton. The primary objective of this study is twofold: (i) to develop and validate an integrated, reproducible computer-vision pipeline for the automated detection and temporal recognition of badminton strokes from broadcast video; and (ii) to use this validated pipeline as a measurement tool for an exploratory, descriptive characterisation of the movement and stroke-sequence patterns associated with scoring versus conceding rallies. We emphasise that the tactical analyses in this work are exploratory and correlational; the Markov-chain and sequential (lag) descriptions are intended to characterise observed patterns rather than to establish causal mechanisms.

Contributions

The main contributions of this work are: (1) an integrated framework that couples self-supervised pretraining (ConvNeXt V2), a Transformer-based temporal module (MotionFormer), and downstream tactical analytics within a single reproducible pipeline; (2) a domain-specific self-supervised pretraining strategy that leverages 139,501 unlabelled broadcast frames and yields a 9.8% relative gain in stroke-recognition mAP@0.5 without additional manual annotation; (3) a biomechanically motivated 10-frame temporal-window design (400 ms) matched to the stroke cycle, together with an annotation-free pipeline for generating temporal training samples; and (4) an exploratory, gender-stratified characterisation of movement and stroke-sequence patterns in elite singles play, reported with effect sizes, 95% confidence intervals, and mixed-effects models. We position BadmintonVision relative to prior badminton pipelines such as ShuttleSet [15] and the TrackNet family of shuttlecock trackers [7, 8, 12]: rather than tracking the shuttle or providing stroke-level labels alone, our framework integrates detection, temporal recognition, and tactical analytics into one workflow. MotionFormer is not a newly designed attention operator, but a lightweight temporal-aggregation module adapted from the standard Transformer encoder [17], applied on top of frozen self-supervised detection features.

Materials and methods

The research workflow

In this study, we collected videos from the BWF World Championships and segmented them into short clips corresponding to individual scoring rallies, recording which player scored at the end of each rally. Expert annotators then labelled a portion of these clips. The overall analytical pipeline proceeded as follows (Fig. 1).

Fig. 1.

The research workflow diagram

First, we trained six YOLO models on the annotated data and selected the best-performing model as the base. To further improve detection performance, we pre-trained this model using SSL, which enhanced its capability to detect targets in complex scenes. Because professional badminton movements are often continuous, we integrated a temporal model to better recognize sequential actions, particularly defensive strokes and jump smashes. Finally, we calculated movement speeds, applied trajectory smoothing, and computed a Markov chain transition matrix to characterize stroke-sequence transitions. Together, these components form an efficient and integrated framework for automated tactical analysis in competitive badminton.

Datasets collection

We gathered official broadcast footage from five BWF World Championships held between 2019 and 2025 (Table 1). All recordings were encoded at 25 frames per second (fps), and this frame rate was consistent across every competition and gender subset listed in Table 1. To ensure consistent camera positioning and elite-level competition, we focused exclusively on singles matches (men’s and women’s) from the quarterfinal stage onwards.

Table 1.

Video data statistics by championship and gender, including the consistent source-video frame rate

Year	Venue	Gender	Rally Clips	Avg. Duration (s)	SD (s)	Frame rate (fps)
2025	Paris	Men	658	12.77	6.2	25
2025	Paris	Women	545	13.14	6.6	25
2023	Copenhagen	Men	642	12.69	5.9	25
2023	Copenhagen	Women	537	13.34	6.4	25
2022	Tokyo	Men	583	13.28	6.1	25
2022	Tokyo	Women	615	13.87	6.8	25
2021	Huelva	Men	569	13.09	6.0	25
2021	Huelva	Women	597	14.65	7.1	25
2019	Basel	Men	520	13.40	5.8	25
2019	Basel	Women	599	14.71	7.0	25
Total	-	-	5,865	13.49	6.4	25 (all)

Videos were segmented into individual rally clips, defined from service contact to point conclusion, yielding 5,865 rally clips in total. The rally-count-weighted mean duration across the ten subsets was 13.49 s; the unweighted mean of the ten subset means was also 13.49 s after rounding. The pooled standard deviation was 6.4 s (per-subset standard deviations 5.8–7.1 s; Table 1), reflecting the wide range of rally lengths in elite singles play. To construct the object detection dataset, we extracted frames at approximately 10-frame intervals, producing 208,891 candidate images. Quality filtering based on Laplacian variance (threshold = 100) removed motion-blurred frames, resulting in 169,501 high-quality images (81.1% retention rate). Of these 169,501 high-quality images, 30,000 were selected for expert annotation to construct Data_L3k, and the remaining 139,501 images were used as the unlabelled corpus for self-supervised pretraining.

For expert annotation, we employed stratified random sampling to select 3,000 images from each of the ten subsets (5 championships × 2 genders), yielding 30,000 images for manual labelling (Data_L3k). Three certified badminton coaches (> 8 years of coaching experience) annotated five object classes targeting the near-court player: Player (bounding box around the near-court player), Forehand stroke (FH; any forehand technique), Backhand stroke (BH; any backhand technique), Serve (service motion, short or long), and Jump Smash (JS; aerial smash with both feet off the ground), as shown in Fig. 2a. The dataset includes left- and right-handed players. Prior to full annotation, inter-rater reliability was established on a 500-image subset, achieving substantial agreement (Cohen’s κ = 0.89). There were 29,256 (97.5%) target frames for Player, 7,438 (24.8%) for FH, 4,056 (13.5%) for BH, 1,089 (3.6%) for Serve, and 1,247 (4.2%) for JS, as shown in Fig. 2b. The target-box counts by competition and gender are shown in Fig. 2c.

Fig. 2.

Summary of data dimension styles and categories (a) Target box category style (b) Category box quantity distribution (c) Distribution of target category box in different time events

Object detection and self-supervised learning

In this study, we explored the application of advanced object detection models for automated tactical analysis in elite badminton. Specifically, we evaluated several variants of the YOLO architecture, including YOLOv8s, YOLOv8m [18], YOLOv10s, YOLOv10m [19], YOLOv12s, and YOLOv12m [20]. Through experiments, we found that the YOLOv12m model outperformed the others in terms of both accuracy and speed, making it the chosen base model for further enhancement.

To improve the object detection capabilities of YOLOv12m, we applied two SSL methods, SimSiam [21] and ConvNeXt V2 [16], to pre-train the backbone network of YOLOv12m. These self-supervised methods allowed YOLOv12m to learn meaningful features from a large amount of unlabelled badminton data, enhancing its performance in complex and dynamic environments, particularly for detecting player movements and shuttlecock trajectories in badminton games.

SimSiam is a self-supervised contrastive learning method, where the main idea is to learn image features by maximizing the similarity between different augmented versions of the same image. First, the original image is augmented to generate two different versions, x₁ and x₂. These two images are then passed through YOLOv12m’s backbone network, which produces feature vectors z₁ and z₂, corresponding to the feature representations of images x₁ and x₂, respectively. The goal of SimSiam is to make these feature vectors as similar as possible by minimizing the negative cosine similarity loss, which is defined as:

1

where z₁ and z₂ denote the feature vectors from two augmented views, and ||·||₂ denotes the Euclidean norm. The loss is therefore the negative cosine similarity between the two L2-normalised feature vectors. Following the standard SimSiam formulation, a stop-gradient operation was applied to one branch. Optimizing this loss encourages the network to learn augmentation-invariant visual representations from badminton frames.

ConvNeXt V2 is a self-supervised method based on masked image modelling, where parts of the image are randomly masked, and the network is trained to reconstruct the masked regions. For our training of YOLOv12m, ConvNeXt V2 was also applied to its backbone network. Specifically, given an image x, a binary mask is used to randomly mask a portion of the image, generating the masked image x⊙M, where ⊙ represents element-wise multiplication. The network then predicts the missing region from the masked image x⊙M and optimizes the model by minimizing the reconstruction error:

2

where is the predicted masked pixel, and x_m is the actual pixel value. This reconstruction loss helps the network learn spatial structures and long-range dependencies in the image, thus improving YOLOv12m’s performance in complex scenarios.

Temporal modelling

While SSL-enhanced YOLOv12m substantially improved single-frame detection, stroke actions inherently span multiple frames. Forehand, Backhand, and Jump Smash recognition remained challenging due to rapid motion blur and pose ambiguity within individual frames. To address this limitation, we developed MotionFormer, a Transformer-based temporal aggregation module that exploits sequential dependencies across consecutive frames.

The key insight of the MotionFormer design is that stroke actions follow predictable biomechanical phases: preparation (body rotation and racket backswing), execution (forward swing and contact), and recovery (follow-through and return to stance). These phases manifest as distinctive motion patterns across 9–14 frames (360–560 ms at 25 fps). By aggregating information from this temporal window, the model can resolve ambiguities that confound single-frame analysis—for instance, distinguishing Forehand from Backhand based on preceding shoulder rotation rather than instantaneous arm position alone.

Temporal training data generation

Because the Data_L3k dataset comprises isolated frames spaced roughly 10 frames apart rather than continuous video sequences, manually re-annotating temporal windows would require prohibitive effort. We therefore developed an automated pipeline that generates temporal training samples without additional annotation (Fig. 3).

Fig. 3.

Schematic diagram of temporal model construction process

For each annotated frame at time t with ground-truth label y_t, we locate its position within the source video and extract a symmetric window of W = 2k + 1 consecutive frames: {I_t−k, …, I_t, …, I_t+k}. The centre frame retains its original annotation, which serves as the supervision signal for the entire window. This approach assumes temporal coherence, that the action occurring at frame t is consistent across the short window duration, a reasonable assumption given typical stroke execution time of 300–500 ms. From the 30,000 annotated images distributed across 5,865 rally clips, we generated training samples through four steps: (1) video-frame mapping, where each annotated image was traced to its source rally clip and frame index using embedded metadata; (2) window extraction, where frames with sufficient temporal context (not within k frames of clip boundaries) had their surrounding windows extracted; (3) sample construction, where each sample comprised the tuple ({I{t − k}, …, I{t + k}}, y_t)—the frame sequence paired with the centre-frame label; and (4) dataset splitting into 80% training (24,000 samples) and 20% validation (6,000 samples), stratified by action class and video source.

Temporal model architecture

Each frame I_i in the temporal window passes through the frozen SSL+YOLO backbone to extract a feature vector:

3

where is the feature dimension. The complete window yields a feature sequence:

4

The transformer encoder processes the feature sequence through the following components (Fig. 3). Learnable position embeddings encode temporal order, enabling the attention mechanism to distinguish frame positions:

5

The model comprises L = 4 identical Transformer encoder layers [17, 22], each containing multihead self-attention (MHSA) and a feed-forward network (FFN) with residual connections and layer normalization:

6

with attention heads, each head computes scaled dot-product attention:

7

where Q, K, V are linear projections, and d_k = D/H = 32 is the per-head dimension. The attention mechanism learns to weight frames containing discriminative motion cues, typically the preparation-to-execution transition, more heavily than static or ambiguous frames. The feed-forward network applies a two-layer MLP with ReLU activation and expansion factor 4:

8

where and . The output representation at the centre frame position is passed through a linear classifier to yield predicted probabilities over C = 5 action classes:

9

Temporal training configuration

The MotionFormer module was configured with feature dimension D = 256, L = 4 Transformer layers, H = 8 attention heads, FFN dimension of 1024, and dropout rate of 0.1. Training employed AdamW optimizer with initial learning rate of 1 × 10^− 4 (cosine decay schedule), weight decay of 0.01, and batch size of 16 for 100 epochs. The model was trained to minimize cross-entropy loss:

10

The trained YOLO backbone weights were frozen [23], serving purely as a feature extractor, while only the newly introduced Transformer encoder and classification head are trained. This strategy offers three advantages: it preserves the motion-discriminative representations learned through SSL; it dramatically reduces trainable parameters (from approximately 25 million to 2 million); and it prevents catastrophic forgetting of single-frame detection capabilities.

Player movement speed calculation

Player movement speed is derived from sequential bounding box centroid positions across consecutive frames (Fig. 4a). For detected bounding boxes at frames and with image-space centroids (ut, vt) and (ut + 1,vt + 1), the real-world displacement is computed following homography transformation [24]:

11

Fig. 4.

Analysis of movement speed, homography transformation, and the Markov-chain method. a Movement speed. b Homography transformation. c Markov chain. Abbreviations: FH, Forehand; BH, Backhand; JS, Jump Smash

where (x, y) denotes court-coordinate positions. Instantaneous speed is then:

12

with Δt = 1/fps = 0.04 s at 25 fps. To reduce measurement noise, we applied temporal smoothing using a 5-frame moving average while preserving rapid directional changes characteristic of badminton footwork.

Homography-based coordinate transformation

Image coordinates were mapped to real-world court coordinates using a perspective transformation matrix computed from manually identified court corner points [25]. Following the official Badminton World Federation court specifications, the singles court measures 13.40 m in length and 5.18 m in width, while the 6.10 m width corresponds to the doubles court; accordingly, we adopted 5.18 m as the singles court width. Player movement was modelled within one half-court, namely the 6.70 m net-to-baseline dimension, corresponding to half of the official 13.40 m singles court length, by the 5.18 m singles width nearest the broadcast camera, so that both players’ trajectories were expressed in a common metric court frame. A separate homography was estimated for each match from its own manually identified court corners, rather than assuming a single constant transformation across matches, to accommodate differences in camera placement and zoom. The homography matrix H satisfies (Fig. 4b):

13

where (u, v) represents image coordinates, (x, y) represents court coordinates in meters, and s is a scale factor. The transformation maps the perspective-distorted broadcast view to a normalized top-down representation of the half-court (6.70 × 5.18 m), enabling metric-scale measurements of player positions and movements. Accordingly, in all subsequent movement and speed equations the coordinates x and y denote these post-homography court positions expressed in metres (not raw image pixels), so that displacements, speeds, and path lengths are reported directly in metric court units.

Markov chain analysis

Stroke-sequence dynamics were modelled as a first-order Markov chain (Fig. 4c), where transition probabilities capture tactical patterns in shot selection [26]. We distinguish this first-order Markov formulation, which characterises one-step transitions between consecutive stroke types, from sequential or lag analysis in observational methodology [27, 28]: lag-sequential analysis tests whether a behaviour follows another at a specified lag more often than expected by chance using adjusted residuals, whereas the Markov matrix summarises the aggregate one-step dependency structure across a rally. To complement the Markov description, we additionally computed lag-1 sequential adjusted residuals between stroke types. For stroke types S ∈ {FH, BH, JS}, the transition probability from one stroke to the next was estimated as:

14

where N_ij counts observed transitions from stroke i to stroke j across all analysed rallies. The transition matrix P reveals tactical signatures: elevated P_FH→JS indicates aggressive finishing patterns, while high P_BH→BH suggests defensive positioning. Separate matrices were computed for scoring and conceding rallies to identify outcome-associated tactical differences.

Trajectory smoothing

Raw trajectory coordinates extracted from frame-by-frame detections contain high frequency noise from detection jitter and player micro-movements. Savitzky-Golay filtering removes this noise while preserving genuine direction changes critical for tactical analysis:

15

where c_j are convolution coefficients derived from local polynomial fitting. We employed window size 2m + 1 = 5 and polynomial order p = 2, balancing noise suppression against temporal resolution. This configuration preserves sharp directional transitions occurring over 3 + frames while attenuating single-frame outliers.

Model training

All models were trained on a workstation equipped with a single NVIDIA GeForce RTX 4090 GPU (24GB VRAM). YOLOv12m was trained using stochastic gradient descent with momentum 0.937, weight decay 0.0005, and initial learning rate 0.01 with cosine annealing schedule over 100 epochs. Batch size was set to 16 with input resolution 640 × 640 pixels. Data augmentation included mosaic augmentation, HSV colour jittering, and random scaling (0.5–1.5×).

Self-supervised pretraining utilized AdamW optimizer with learning rate 0.0001, weight decay 0.05, and batch size 256 for 200 epochs. SimSiam employed symmetric loss with predictor network, while ConvNeXt V2 used 75% masking ratio following MAE protocol.

To avoid information leakage from temporally adjacent frames, all data partitioning was performed at the match level rather than at the frame level: every frame, rally clip and temporal window originating from a given match was assigned in its entirety to a single split, so that no rally or video was shared across the training, validation and test sets. The 139,501 unlabelled images used for self-supervised pretraining were drawn from matches that are disjoint, at the clip level, from the rallies used for downstream evaluation, ensuring that the pretraining corpus did not overlap with the test material. The verification results are reported at the end of Sect."Temporal modelling performance".

Evaluation metrics and statistical analysis

Detection performance was evaluated using Average Precision (AP) at an intersection-over-union (IoU) threshold of 0.5, computed as the area under the precision–recall curve. Mean AP (mAP@0.5) aggregates performance across all five object classes. Detection inference used a confidence threshold of 0.25 and non-maximum suppression at an IoU threshold of 0.7. For trajectory analysis, we computed path efficiency, direction changes, centre-court occupancy, and base recovery time. Path efficiency was the ratio of straight-line displacement to total distance travelled. Direction changes were the number of successive reversals in smoothed movement-vector orientation exceeding 90° within a rally. Tactical diversity was quantified in the Methods as Shannon entropy, H = −Σ p_i log₂(p_i), where p_i is the proportion of strokes of type i within a rally; higher H indicates a more even distribution of stroke types. Scoring and conceding conditions were compared using two-sided independent-samples t-tests, with Cohen’s d and 95% confidence intervals reported for effect size and mean difference. The significance threshold was α = 0.05. Because rallies were nested within games, we also fitted linear mixed-effects models with outcome condition as a fixed effect and a random intercept for game; fixed-effect estimates (β) and 95% confidence intervals were reported in native metric units. Markov transition probabilities and lag-1 adjusted residuals were treated descriptively because the gender-stratified tactical sequence samples were exploratory and contained few games.

Tactical-analysis sampling and outcome assignment

For movement analysis, the sampling unit was an individual rally clip. We analysed 147 rally clips drawn from four games within two elite singles matches: two men’s games (n = 74 rallies) and two women’s games (n = 73 rallies). The pooled outcome groups contained 74 scoring rallies and 73 conceding rallies. Only the near-court player was measured in any given rally. Because players changed court ends between games, both participants could be represented across the sampled games, but they were never measured simultaneously within the same rally. Each rally contributed once—and only once—to the analysis: it was classified as scoring when the near-court player won the rally and as conceding when the near-court player lost it.

For stroke-sequence analysis, the sampling unit was likewise an individual rally clip. We analysed 70 rallies from two games (one men’s game, n = 34; one women’s game, n = 36). For phase analysis, strokes were assigned by normalized ordinal position within each rally: the opening phase comprised the first 20% of strokes, the middle phase the central 60%, and the ending phase the final 20%. This proportional definition, rather than an absolute stroke-count cutoff, allowed rallies of different lengths to be compared.

Results and analysis

YOLO and SSL performance analysis

The YOLO benchmark results are reported in Fig. 5; Table 2. Player detection achieved AP@0.5 values of 0.912–0.947 across the six models, Serve achieved 0.824–0.871, Forehand achieved 0.698–0.763, Backhand achieved 0.672–0.742, and Jump Smash achieved 0.638–0.705 (Fig. 5a–c).

Fig. 5.

YOLO baseline and SSL performance comparison. a Class-wise AP@0.5 comparison across six YOLO models. b Overall mAP@0.5 and mAP@0.5:0.95 comparison. c Speed-accuracy trade-off scatter plot. d Class-wise AP@0.5 comparison among the three selected methods. e Relative improvement percentage. f Radar-chart comparison between the baseline and ConvNeXt V2

Table 2.

Baseline YOLO model performance, bold font indicates that the model index is the best compared with other models

Models	mAP@0.5	Player	Forehand	Backhand	Serve	Jump Smash
YOLOv8s	0.749	0.912	0.698	0.672	0.824	0.638
YOLOv8m	0.772	0.928	0.724	0.701	0.842	0.667
YOLOv10s	0.762	0.921	0.712	0.689	0.836	0.654
YOLOv10m	0.787	0.938	0.741	0.718	0.856	0.682
YOLOv12s	0.777	0.932	0.729	0.706	0.849	0.671
YOLOv12m	0.806	0.947	0.763	0.742	0.871	0.705

Among baseline models, YOLOv12m achieved the highest overall performance (mAP@0.5: 0.806), demonstrating improved feature extraction for fine-grained action discrimination compared to earlier architectures. However, the persistent accuracy gap in action recognition motivated our investigation of self-supervised pretraining strategies. Therefore, we leveraged the 139,501 unlabelled images remaining after expert annotation to pretrain visual encoders using two self-supervised approaches: SimSiam and ConvNeXt V2. These pretrained representations were subsequently integrated with YOLOv12m through feature fusion (Fig. 5d). Here, feature fusion denotes that the self-supervised encoder was used to initialise and augment the YOLOv12m backbone rather than to replace it: the SSL-pretrained ConvNeXt V2 weights were transferred into the detector backbone, and the resulting backbone features were propagated through the standard YOLO feature-pyramid neck, so that the detection head operated on representations enriched by self-supervised pretraining. SimSiam pretraining yielded modest improvements (mAP@0.5: 0.840, + 4.2% relative gain), with incremental gains distributed across all classes. ConvNeXt V2, employing masked autoencoding with a hierarchical ConvNet architecture, demonstrated substantially greater efficacy (mAP@0.5: 0.885, + 9.8% relative gain, Fig. 5e).

ConvNeXt V2 pretraining improved Forehand AP@0.5 from 0.763 to 0.879, Backhand from 0.742 to 0.856, and Jump Smash from 0.705 to 0.812 (Fig. 5f).

Temporal modelling performance

The temporal-model test results are shown in Fig. 6, with complete class-wise values provided in Supplementary Table S1. At IoU = 0.5, mAP increased from 0.885 for the single-frame baseline to 0.902 with temporal smoothing, 0.922 with the 5-frame model, and 0.941 with the 10-frame model, before reaching 0.934 with the 15-frame model (Fig. 6a–c). Relative to the single-frame baseline, the 10-frame MotionFormer model provided a 6.3% gain in mAP@0.5. Class-wise values are reported in Supplementary Table S1.

Fig. 6.

Temporal modelling: (a) class AP; (b) window comparison; (c) mAP; (d) baseline improvement; (e) biomechanical windows; (f) split comparison; and (g) coherence audit. Only panel d shows Δ between the 10-frame and 15-frame conditions because panels e and f are not temporal-window comparisons. Panel e reports 200 ms, 400 ms, and 600 ms. FH, Forehand; BH, Backhand; JS, Jump Smash

Relative to the single-frame model, the 10-frame input increased AP@0.5 by 1.2% points for Player, 6.2 for FH, 7.1 for BH, 1.6 for Serve, and 12.0 for JS (Fig. 6d). The corresponding 15-frame values were 0.981, 0.932, 0.918, 0.921, and 0.919, respectively, compared with 0.983, 0.941, 0.927, 0.924, and 0.932 for the 10-frame condition. At 25 fps, the tested 5-frame, 10-frame, and 15-frame windows corresponded to 200 ms, 400 ms, and 600 ms, respectively (Fig. 6e).

Because adjacent broadcast frames are highly correlated, a frame-level random split inflates apparent performance; we therefore report the match-level mAP@0.5 of 0.941 as our primary, conservative result, and provide the corresponding frame-level value (0.956) only to quantify the inflation that match-level splitting avoids (Fig. 6f). To examine the temporal-coherence assumption underlying the 10-frame windows, we audited the label consistency of window-edge frames on 300 randomly sampled windows: 290 of 300 (96.7%) were temporally consistent, and re-evaluating after excluding ambiguous window-edge frames changed mAP@0.5 only marginally (0.941 versus 0.943), indicating that window-edge label noise does not materially affect the reported results (Fig. 6g).

Player trajectory and movement analysis

Player-trajectory results are presented in Fig. 7, with pooled scoring–conceding estimates provided in Supplementary Table S2. The sampling procedure and the one-rally/one-condition assignment are specified in Sect. "Tactical-analysis sampling and outcome assignment".

Fig. 7.

Player trajectory and movement analysis. The labels Scoring and Conceding and the ordinate order are used consistently across panels. a Average speed by player and outcome condition. b Path efficiency by outcome condition. c Centre-court occupancy by outcome condition. d Heat map of men’s singles positions (score-loss differential), with the central area dominating in scoring. e Heat map of women’s singles positions (score-loss differential) also exhibiting a central dominance pattern. f Cohen’s d estimates with 95% confidence intervals. g Mixed-effects fixed-effect estimates (β) in native units with 95% confidence intervals from models with a random intercept for game

Scoring rallies had an average speed of 2.42 ± 0.30 m/s compared with 2.32 ± 0.30 m/s in conceding rallies (p = 0.056, d = 0.32), so this difference was non-significant (Fig. 7a). Scoring rallies showed greater path efficiency (0.74 ± 0.10 vs. 0.57 ± 0.10; d = 1.77) (Fig. 7b), greater centre-court occupancy (45.0 ± 7.6% vs. 36.8 ± 8.8%; d = 1.00) (Fig. 7c-e), shorter base recovery time (0.85 ± 0.20 s vs. 1.15 ± 0.19 s; d = − 1.56), and more direction changes (3.84 ± 1.73 vs. 2.68 ± 1.38; d = 0.74) (Fig. 7f; Supplementary Table S2). The mixed-effects estimates with a random intercept for game are reported in Fig. 7g.

Gender-stratified results are presented in Fig. 7a–c. Men had higher average movement speeds than women (2.48 vs. 2.25 m/s; p < 0.01). Across the pooled sample, scoring rallies had greater path efficiency (0.74 vs. 0.57; mixed-effects β = +0.17) and greater centre-court occupancy (45.0% vs. 36.8%; mixed-effects β = +8.20% points) than conceding rallies (Fig. 7g).

Stroke pattern and scoring correlation

Using the sampling and definitions specified in Sect."Evaluation metrics and statistical analysis" and "Tactical-analysis sampling and outcome assignment, we analysed stroke sequences from 70 rallies in one men’s game (n = 34) and one women’s game (n = 36). In men’s scoring rallies, the FH→JS transition probability was 37%. In women’s scoring rallies, the FH→FH transition probability was 43%. In conceding rallies, BH→BH transitions were 62% for men and 57% for women (Fig. 8a).

Fig. 8.

Stroke-sequence and tactical-signature results: (a) Markov transition matrices; (b) stroke composition during opening (first 20%), middle (central 60%), and ending (final 20%) phases; (c) terminal three-stroke patterns; (d) tactical fingerprint using mean rally length rather than an undefined short-rally category; (e) point-win probability by fully named terminating stroke, with S/C abbreviations removed; (f) rally length versus shot-sequence entropy; and (g) lag-1 adjusted residuals. FH, Forehand; BH, Backhand; JS, Jump Smash

Stroke composition varied across the proportionally defined opening, middle, and ending phases (Fig. 8b). In the ending phase, Jump Smash represented 56.2% of men’s strokes and 31.2% of women’s strokes. Terminal three-stroke patterns were also gender-stratified: FH→FH→JS occurred in 7.7% of men’s rally endings, whereas FH→FH→FH occurred in 9.5% of women’s rally endings (Fig. 8c).

The exploratory tactical fingerprint showed a higher Jump Smash finish rate for men than women (69% vs. 33%) and a shorter mean rally length (9.2 ± 2.4 vs. 10.9 ± 2.3 strokes) (Fig. 8d). In the terminating-stroke analysis, pressured Backhand endings were associated with point loss in 76% of men’s rallies and 73% of women’s rallies (Fig. 8e). Shot-sequence entropy increased with rally length in both gender-stratified samples (Fig. 8f). Lag-1 adjusted residuals are reported in Fig. 8g. Because each sample comprised only one game, these relationships are descriptive.

Discussion

Automated tactical analysis in racket sports has remained challenging due to the rapid, transient nature of stroke actions and the complexity of extracting meaningful performance metrics from broadcast video. Here, we present BadmintonVision, an integrated deep learning framework that addresses these challenges through three synergistic innovations: self-supervised pretraining on domain-specific unlabelled data, temporal modeling matched to stroke biomechanics, and downstream tactical analysis pipelines. Applied to elite badminton footage from BWF World Championships, our framework achieved 94.1% detection accuracy (mAP@0.5) for stroke recognition and enabled quantitative characterization of movement patterns and tactical signatures associated with rally outcomes.

The largest SSL-associated gains occurred in Forehand, Backhand, and Jump Smash, the three classes with the lowest baseline AP. Rapid execution, transient poses, and motion blur provide plausible explanations for their lower single-frame performance. Domain-specific self-supervised pretraining may improve representations of body posture and racket orientation, but direct feature analysis is required to establish the mechanism [29].

The 10-frame temporal window provided the highest observed mAP. At 25 fps, its 400 ms duration is consistent with the 300–500 ms stroke cycle described in the Methods, whereas the 200 ms 5-frame window may omit preparation or follow-through cues and the 600 ms 15-frame window may include recovery or the next action. This interpretation is consistent with the observed performance pattern but would require direct attention-map or feature-attribution analysis for confirmation.

The movement findings align with coaching emphasis on central positioning and efficient recovery, but the observational design supports association rather than causation. A plausible interpretation is that skilled players reach tactically advantageous positions more efficiently; alternatively, opponent quality or game context may produce the same pattern. The stroke-transition results likewise identify candidate tactical signatures for confirmation rather than prescriptive rules.

The exploratory gender-stratified patterns suggest that successful sequences may be achieved through different combinations of power-oriented and sustained play. Because the sequence analysis included only one game per gender, these patterns should guide hypotheses for larger, player-balanced studies rather than gender-specific training recommendations.

From a methods perspective, BadmintonVision provides a flexible framework for sports analysis using computer vision. Its modular design, comprising detection, temporal modelling, coordinate transformation, and tactical analysis, can be applied to other racket sports like tennis, squash, and table tennis, or even to team sports with adjustments. By using SSL, we address the challenge of limited expert-annotated data. Our approach takes advantage of abundant unlabelled broadcast footage, reducing the need for manual annotation while improving model performance.

Several practical uses come from this work. A real‑time version could give coaches immediate feedback on tactical patterns during practice or between games. Long‑term analysis across tournaments could track how a player’s tactics change or find new strategy trends at the top level. The framework can also produce objective scouting reports that measure opponent tendencies in stroke choice and movement. In addition, the annotated dataset and trained models provide resources for the research community to build and evaluate future badminton analytics methods.

This study has seven principal limitations. (1) We analysed singles play only; extension to doubles and lower-level competition requires validation. (2) Only the near-court player was observed in each rally, so players were never measured simultaneously; each rally entered either the scoring or conceding condition once. (3) The proof-of-concept tactical samples were limited to 147 rally clips from four games for movement analysis and 70 rally clips from two games for stroke-sequence analysis. The absolute magnitudes of Cohen’s d ranged from 0.32 to 1.77, and their confidence intervals should be interpreted in light of the small number of games. (4) Classification was limited to five object classes, including four stroke actions; finer stroke labels require additional annotation. (5) The current system operates on recorded video; real-time deployment requires streaming optimization. (6) All source matches were elite BWF World Championships contests from the quarterfinal stage onwards (Table 1), which limits generalizability to recreational, amateur, or intermediate play. (7) The correlational design precludes causal inference; for example, efficient movement may reflect player skill, opponent quality, or game context rather than directly determining rally outcome.

Conclusions

In this study, we presented BadmintonVision, a deep learning framework for automated tactical analysis in elite badminton. The framework combines self-supervised pretraining, temporal modelling, and tactical analysis to address the key limitations of traditional manual video review. By pretraining the detection model on 139,501 unlabelled images using ConvNeXt V2, we reduced the need for large-scale labelled data while providing a 9.8% relative gain in stroke recognition. The proposed MotionFormer module processes 10-frame windows to capture complete stroke cycles and provided an additional 6.3% relative gain, leading to an overall detection accuracy of 94.1% mAP@0.5. Based on a dataset of 30,000 expert-annotated images from BWF World Championships (2019–2025), our movement analysis showed that scoring rallies are associated with higher path efficiency and shorter recovery times, with effect-size absolute magnitudes ranging from 0.32 to 1.77. These associations, while consistent with coaching intuition, require replication with larger and more diverse samples before drawing strong conclusions about their practical significance for training interventions. In addition, the Markov chain analysis of stroke sequences suggested preliminary, gender-associated tactical patterns; because only a single game was analysed per gender, these patterns require confirmation in larger and more balanced samples before they can reliably inform match preparation and strategy design. However, this study has some limitations: the current framework covers only five action classes and has been assessed mainly on professional-level matches, so its ability to handle more complex actions and lower-level competitions still needs further validation. Future work will focus on expanding the action categories, incorporating real-time analysis capabilities, and assessing the framework across different skill levels to improve its practical value for coaching and performance research.

Authors'' contributions

YZ conceived and supervised the study. YL developed the methodology, implemented the software, performed the formal analysis, curated the data, prepared the visualizations, and drafted the manuscript. GZ contributed to the investigation and provided resources. YZ reviewed and edited the manuscript and administered the project. All authors read and approved the final manuscript.

Funding

No funding was accessed for this research.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author upon reasonable request following publication of the article. To support reproducibility, the model training and analysis code, the configuration files, and the full annotation guidelines are publicly available at https://github.com/liangyongjie51/BadmintonVision, and the self-supervised and detection pretrained weights are released through the same repository.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki and adhered to the ethical guidelines of the International Committee of Medical Journal Editors (ICMJE). The study exclusively involved computational analysis of publicly available broadcast footage from BWF World Championships (2019–2025), which is freely accessible through official media channels. No human participants were recruited, no personal or identifiable data were collected, no biological samples were obtained, and no interventions were performed. All analysed footage depicted professional athletes competing in publicly televised events where no reasonable expectation of privacy exists. In accordance with national legislation and the institutional requirements of Guizhou University of Finance and Economics, formal ethical review and approval were waived, as the research did not involve direct interaction with human subjects or the processing of personally identifiable information. Written informed consent was not applicable, as the study relied solely on secondary analysis of publicly broadcast sports footage.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.