Making the World’s Fastest Racket Sport even Better: A Systematic Review of Artificial Intelligence-based Objective Player Performance Assessment in Badminton

Making the World’s Fastest Racket Sport even Better: A Systematic Review of Artificial Intelligence-based Objective Player Performance Assessment in Badminton | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Systematic Review Making the World’s Fastest Racket Sport even Better: A Systematic Review of Artificial Intelligence-based Objective Player Performance Assessment in Badminton Christian Arthur, Raymond Chiong, Fareed Ud Din, Farshid Hajati, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9105158/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Artificial intelligence (AI) is increasingly promoted as a means of providing objective player performance assessment in badminton. Compared to other sports, however, the supporting evidence remains fragmented. A systematic review based on 51 studies that satisfied the established quality and eligibility criteria from three major databases (covering the period 2018 to the end of 2025) reveals four dominating methodological schools: computer vision stroke tracking, movement-pattern recognition, spatio-temporal analysis of rally sequences, and multi-modal frameworks that integrate several data streams. Although many studies report high classification or prediction accuracy, only a small proportion of them employ shared validation datasets or evaluate repeatability across testing sessions, which limits the generalisability of their findings. Common shortcomings include small or imbalanced data samples, weak alignment with established sport-science theory, substantial computational requirements, and participant pools drawn largely from elite athletes in a single geographic region. Recent work has begun integrating explainable AI with retrieval-augmented generation (RAG) and large language model (LLM) frameworks to provide grounded, query-responsive feedback that links visual detections and performance metrics to structured match evidence. Future research should focus on larger and more diverse datasets, alignment with skill development models, transparent output formats, and validation across competitive levels and contexts with these state-of-the-art explainable AI-based RAG or LLM frameworks. Artificial intelligence badminton computer vision player assessment sports analytics Figures Figure 1 Figure 2 Figure 3 1. Introduction Since its inclusion in the Olympic Games in 1992, badminton has developed into a highly competitive racquet sport with a well-established professional structure [ 1 ]. The sport is widely recognised as the fastest racquet sport, characterised by extremely rapid rallies and shuttle speeds that demand quick reactions, precise technical execution, and continuous tactical adaptation. It is played recreationally by an estimated 200 million people worldwide, with more than 7,000 athletes competing annually in hundreds of national and international events [ 2 ]. Traditionally, coaches, selection committees, and technical administrators have based badminton player assessment on subjective evaluation [ 3 , 4 ]. While capitalising on the domain expertise of the practitioner, this approach can lead to inconsistent and biased approaches to player development, competitive structuring, and participant satisfaction [ 5 , 6 ]. The subjective nature of traditional assessment models has spurred increasing interest in more objective, data-driven approaches to player evaluation [ 7 ]. Compared to other sports, badminton is a complex one to assess because of its multiple technical requirements, fast pace of play, and the presence of multiple skill components [ 8 , 9 ]. To be proficient in multiple domains, players must have demonstrated proficiency in stroke techniques, court movement, tactical decision-making, and competitive temperament [ 10 , 11 ]. However, these skills are integrated at speeds much faster than other racket sports [ 12 ]. In most countries worldwide, club-level badminton associations use classification systems to divide players into different skill categories to help establish appropriate competitive structures and training organisation [ 13 ]. Despite this, the subjectivity of assessment often undermines the effectiveness of these systems, resulting in inconsistent assessment, player disputes, and the ability to manipulate for competitive advantage [ 14 – 16 ]. The integration of artificial intelligence (AI) technologies provides promising pathways to address these challenges by providing a systematic, data-driven multi-factor analysis of player performance [ 17 ]. With recent advances in computer vision, machine learning, and data analytics, opportunities exist for more objective assessment frameworks [ 18 ]. However, such systems are theoretically incomplete, methodologically unexplored, and empirically untested [ 19 ]. This review studies AI approaches for badminton player performance assessment. It evaluates their reported accuracy and practical utility, identifies recurring methodological limitations, and highlights underexplored aspects of performance data. By charting prevailing trends and research gaps, we seek to steer future investigations towards rigorous, generalisable, and practitioner-oriented AI frameworks for objective skill assessment in badminton. To the best of our knowledge, this is the first such review on AI-based badminton player performance assessment in the relevant literature. 2. Methods This study follows a systematic literature review methodology guided by the PRISMA framework [20]. The review protocol adapts the procedures outlined by Helbach et al. [21], beginning with the definition of the study’s scope and objectives. Specifically, the review addresses three research questions (RQs): RQ1. Which AI approaches are currently applied to badminton player assessment, and what evidence supports their effectiveness? RQ2. What theoretical, methodological, and technical constraints are most frequently reported across these studies? RQ3. How are AI assessment results communicated to ensure they are clear and trusted by coaches and athletes? These questions structure the subsequent stages of database searching, study screening, data extraction, and synthesis. 2.1. Search Strategy The keywords , presented in Table I, were selected to reflect three core themes: the sport context (e.g., “badminton”, “racket sport”), type of analysis (e.g., “pose estimation”, “performance assessment”), and computational approaches (e.g., “machine learning”, “computer vision”). This structure helped capture a comprehensive yet targeted set of publications relevant to the review’s objectives. Table I Final search queries used across selected academic databases. Database Final query Web of Science Core Collection TS=((badminton OR "racket sport*" OR "racquet sport*") AND (video analysis OR motion tracking OR pose estimation OR action recognition OR movement analysis OR "performance assessment") AND (artificial intelligence OR machine learning OR deep learning OR neural network* OR computer vision)) Scopus (badminton OR "racket sport*" OR "racquet sport*") AND ("video analysis" OR "motion tracking" OR "pose estimation" OR "action recognition" OR "movement analysis" OR "performance assessment") AND ("artificial intelligence" OR "machine learning" OR "deep learning" OR "neural network*" OR "computer vision") IEEE Xplore (("badminton" OR "racket sports" OR "racquet sports") AND (("artificial intelligence" OR "machine learning" OR "deep learning" OR "computer vision" OR "neural networks")) AND (("player assessment" OR "skill classification" OR "performance metrics" OR "technical analysis" OR "sports analytics"))) 2.2. Eligibility Criteria and Study Selection The initial screening was done by the first and fourth authors. Once the screening was completed, all articles that were not related to the topic were excluded from further consideration. For articles lacking a clear title and abstract, full-text screenings evaluated the content to determine whether they should be included. Decisions about which articles to include were made after careful full-text evaluations, with all authors in full agreement. Figure 1 illustrates the data extraction process, providing a clear overview of how the 356 records retrieved from the academic databases were processed to result in the 51 studies that are included in the data synthesis and analysis of this review. Eligibility criteria were established to ensure the relevance of the review to AI-based badminton player assessment. Included studies were peer-reviewed journal articles or conference proceedings published between January 2018 and December 2025 (note: a few early-access articles dated 2026 were also inlcuded) in English. Exclusions applied to abstract-only entries, non-English publications, and studies before 2018. Articles had to focus on AI applications for player assessment, skill evaluation, or performance classification, with methodological clarity and empirical implementation. Studies on tactical analysis, match prediction, or non-AI approaches were excluded unless they addressed player-level assessment. Relevant literature was identified through systematic searches in IEEE Xplore, Scopus, and Web of Science. 3. Results 3.1. Overview of AI Methods Applied to Badminton Player Assessment (RQ 1) 3.1.1. Study Characteristics Figure 2 illustrates the yearly publication trend on AI applications for badminton player assessment. Research output was low from 2018 to 2020, with one publication per year. A clear increase can be seen in 2022, rising to ten publications and remaining similar in 2023 and 2024. The number of studies then increased sharply to nineteen in 2025, indicating accelerated research activities in the field. This trajectory indicates growing scholarly interest coinciding with the maturation of deep-learning toolkits that lower technical barriers to entry. 3.1.2. Methodological Approaches Table II. Methodological approaches employed in AI-based badminton player assessment studies. No. Study Study Focus Player Sample Data Source(s) AI Methodology Applied Performance Metrics Reported Assessment Framework Orientation Validation Approach 1 [22] Stroke-level dataset for tactical performance prediction Elite athletes Video Footage Transformers, Bidirectional-Gated Recurrent Unit (GRU), Graph Models Area Under the Curve (AUC), Accuracy Tactical; Post-hoc Task Benchmarks 2 [5] Automatic point and stroke annotation Elite athletes Video Footage Convolutional neural network (CNN) and Support Vector Machine (SVM) Accuracy, Mean Average [email protected] , Edit Score Technical and Tactical; Post-hoc Split-based Evaluation 3 [6] Profiling performance via score progression Elite athletes Match Data SVM, RF, K-means R², Accuracy, Std. Dev Technical; Post-hoc Train/Test Spl 4 [7] Prediction and tactical decision analysis in women’s singles Elite athletes Video Footage DT, RF, XGBoost, SVM Accuracy (up to 87.5%) Technical and Tactical; Post-hoc Cross-validation (90/10 split) 5 [8] Closed-loop AI for imitation, simulation, and strategy optimisation Elite athletes Video Footage (ShuttleSet, annotated datasets) RallyNet, Simulation Environment Loss metrices, Win Rate Tactical; Real-time and Post-hoc Benchmarking vs baselines (ShuttleNet and DyMF) in simulated matches 6 [9] Behaviour analysis from broadcast videos with visual analytics Elite athletes Video Footage You Only Look Once (YOLO)v3 Model, OpenPose, ResNet-18, LSTM Accuracy (97%), Precision, Recall, F1 (hit-frame detection) Tactical; Real-time and Post-hoc Split-based Evaluation (train/val/test; semi-supervised learning) 7 [10] Accuracy assessment of forehand smash biomechanics Collegiate athletes Video Footage (front-view rallies, Mediapipe extraction) Skeletal Pose Estimation (MediaPipe), Dynamic Model, RF, Ridge Accuracy (up to 97.4%), Precision, Recall, F1 Technical; Post-hoc Train/Test Split (70/30), Cross-validation, Expert validation 8 [11] Deep learning-based recognition of badminton actions Recreational athletes Hybrid Approach Long Short-Term Memory (LSTM), Continuous Learning Accuracy (63%, 84%, 92% across datasets) Technical; Post-hoc Benchmark on HKSR and NTU skeleton datasets 9 [12] Detection and classification of six-corner footwork Recreational athletes Video Footage YOLOv8, YOLOv9 Mean average precision (0.633 vs 0.605), Precision, Recall, F1 Technical; Real-time Train/Val/Test Split (80/10/10); Model comparison 10 [13] Shot and strategy recognition using wearable sensors Elite and semi-professional athletes Wearable Sensors CNN, LSTM Shot Accuracy (90.9%), Strategy Accuracy (80%), Precision, Recall, F1 Technical and Tactical; Post-hoc Train/Val/Test Split (64/20/16), Class weighting 11 [17] Action recognition using spatio-temporal skeleton features Youth athletes Video Footage (3D skeletal data via MediaPipe) Weighted Ensemble Machine Learning Models Accuracy (95.38%), Precision, Recall, F1 Technical; Post-hoc 5-fold Cross-validation 12 [18] Detection and classification of six badminton strokes Elite athletes Video Footage (Kaggle badminton_stroke_video, 3000+ annotated images) YOLO-HGNet mAP (96.1%), Accuracy (95.4%), Precision, Recall, F1 Technical; Real-time and Post-hoc Benchmark vs YOLOv5, OpenPose, CatBoost, XGBoost 13 [23] Footwork recognition and 3D trajectory extraction Sub-elite athletes Video Footage (binocular cameras) Faster R-CNN Binocular Positioning Shoe localisation accuracy (97.2%) and Positioning error (0.129 m) Technical; Real-time and Post-hoc Train/Val/Test Split (90/10), Trajectory validation 14 [24] Player evaluation using DRL with technical and tactical contexts Elite athletes Video Footage (BWF match videos, 2018–2020) DRL (LSTM-based Q-function), AlphaPose, TrackNet Action Value (Q-values), Correlation with Score/Rank Technical and Tactical; Post-hoc Baseline comparisons, Spearman Correlation 15 [25] Shot prediction and classification from broadcast videos Recreational athletes Video Footage (26 broadcast videos, Shuttlecock Trajectory Dataset) YOLOv5, DeepSORT, CNN MOTA (73%), HOTA (72%), IDF1 (73%) Technical; Post-hoc Benchmark with trajectory metrics 16 [26] Classification of badminton shots (lob, smash, net) Elite athletes Video Footage (BWF Men’s Singles matches) Keras-Mediapipe, YOLO-NAS (Neural Architecture Search) Accuracy (73.5%), Precision, Recall, F1, AUC Technical; Post-hoc Train/Val/Test Split (70/10/20), Confusion Matrix 17 [27] Predicting shot accuracy using Quiet Eye and biomechanics Elite, Intermediate, and Novice athletes Hybrid Approach (Eye-tracking + Motion Capture) Neural Network, SHapley Additive exPlanations (SHAP) Accuracy (85.7%), Precision (88.3%), Recall (83.1%), F1 (0.856), AUC (0.823) Technical; Real-time and Post-hoc Train/Val/Test Split (70/15/15), Ablation Study 18 [28] Coaching framework for stroke, stance, and scoring Professional, Intermediate, and Novice athletes Wearable Sensors (IMU: wrist, palm, left leg, right leg) CNN (stroke classification, score regression), kNN (stance error estimation) Stroke Accuracy (89.1%), Performance R² (0.888), Success Rate (%) Technical; Post-hoc Cross-person validation, Train/Val/Test Split (60/20/20), Baseline comparisons 19 [29] Shot influence prediction using long/short-term dependencies Elite athletes Video Footage (43,838 shots, 4,350 rallies, 75 matches, 2018–2021) CNN (short-term), Transformer (long-term) AUC (0.871), Brier Score (0.143), Efficiency gains Tactical; Post-hoc 5-fold Cross-validation 20 [30] Smash detection using optimised deep learning model Elite athletes Video Footage (YouTube broadcast matches) Residual-Shuffle Net + Upgraded Pufferfish Optimiser (UPO) Accuracy (96.4%), Precision (0.95), Recall (0.87), F1 (0.87) Technical; Post-hoc Train/Test Split (85/15), Baseline comparisons (AlexNet, GoogleNet, ResNet-18, I3D) 21 [31] Pose estimation with improved YOLOv8-Pose and attention Recreational athletes Video Footage (xBHPE dataset, Kinect v2, 4000 samples) YOLOv8-Pose + Local Attention Mean squared error (12.72), Percentage of keypoints @0.2 (0.7793), Frame per second (66.4) Technical; Real-time and Post-hoc Benchmark vs HigherHRNet, BlazePose, LitePose 22 [32] Biomechanics, injury prevention, and tactical analysis Elite athletes Hybrid Approach (Video, Wearable Sensors, Historical Match Data) CNN (pose estimation) CV Models (93% accuracy), Injury Prediction (90%), RL Defence Gain (20%) Technical and Tactical; Real-time and Post-hoc Cross-validation, Baseline comparisons, Empirical training tests 23 [33] Automated detection of badminton players from broadcast videos Elite athletes Video Footage (YouTube broadcast matches: 2011 All England, 2012 Olympics, 2017 Asia Championship) Faster R-CNN Average Precision (PR curve-based) Technical; Post-hoc Case-based training/testing on single and combined videos 24 [34] Autonomous motion tracking of squash players Elite athletes Video Footage (broadcast squash matches, PSA Canary Wharf Classic 2013) Multi-person Pose Estimation (CNN), R² (0.99), Error in Distance (3.73%), Speed Error (1.26%) Technical and Tactical; Post-hoc Benchmark vs manual tracking, Filtering vs Unfiltered data 25 [35] 3D shuttle trajectory reconstruction from monocular video Elite athletes Video Footage (TrackNetV2 dataset + 40 YouTube matches) Graph-based court detection, U-Net, GRU Hit Detection (94.6% F1), 3D Reconstruction Error (~8 cm synthetic) Technical and Tactical; Post-hoc Benchmarks on TrackNetV2 Synthetic Trajectory Evaluation 26 [36] Survey of video action recognition in sports (datasets, methods, applications) Multiple sports (team & individual) Video Footage (SoccerNet, Badminton Olympic, Diving48, FineGym, TTStroke-21, etc.) 2D/3D CNN, Two-stream, Transformer, Skeleton-based - Technical and Tactical; Post-hoc Comparative analysis across datasets and models 27 [37] Court line extraction from broadcast badminton videos Elite athletes Video Footage (BadmintonWorld.tv tournament videos) Horizontal Line Projection + K-means, Morphological Corner Detection Pixel Error (0–3 px), Frame Rate (33.3 FPS) Technical; Real-time and Post-hoc Comparative benchmark vs Hough Transform methods 28 [38] Motion trajectory tracking with improved KCF and depth fusion Elite athletes Video Footage (binocular cameras, badminton arm strokes) ViBe (background subtraction), KCF + Depth Fusion Tracking Success Rate (98%), Trajectory Error (0.365), Speed Error (0.116) Technical; Real-time and Post-hoc Benchmark vs Kalman Filter & Original KCF 29 [39] Stroke recognition in badminton using TSN with attention Elite athletes Video Footage (badminton stroke videos) Temporal Segment Network (TSN) + Lightweight Attention Recall (91.2%), Accuracy (91.6%) Technical; Post-hoc Train/Test Split 30 [40] Shot refinement by combining shuttle tracking and hit detection Elite athletes Video Footage (BWF matches, 32 + 354 matches, 8,975 + 98,675 frames) TrackNet (shuttle), YOLOv7 (swing), DensePose (pose), Shot Refinement Algorithm (SRA), Shot Detection: Precision 0.897, Recall 0.913, F1 0.905; Shot Classification: Accuracy 72.1% Technical and Tactical; Post-hoc Benchmark vs TrackNet; Train/Val/Test split; t-IoU evaluation 31 [41] Shuttlecock detection with lightweight small-object detector Elite athletes Video Footage (custom shuttlecock dataset, 6113 train / 1586 test images) YOLO-inspired CNN Mean Average Precision (98% shuttlecock), FPS (30 on Jetson Nano) Technical; Real-time Benchmark vs YOLOv3/v4/v5, v3-tiny, v4-tiny on public and shuttle datasets 32 [42] Rally outcome prediction using stroke sequences Elite athletes Video Footage (London 2012 Olympics, 10 annotated players, 498 rallies) ResNet-18, Bidirectional LSTM, Faster CNN Accuracy (0.70), Brier Score (0.30) Tactical; Post-hoc Train/Test Split (398/100 rallies) 33 [43] Stroke detection using wearable wrist device Recreational athletes Wearable Sensors (Arduino Nano 33 BLE, accelerometer, gyroscope) CNN (embedded), Motion segmentation Accuracy (100% in small dataset), Confusion Matrix Technical; Real-time Prototype testing on limited stroke dataset 34 [44] Optimisation of doubles positioning and movement Elite athletes Match Data (2019 BWF Super Series finals, 10 matches) Numerical Modeling (7 nonlinear equations, MATLAB fsolve) Optimal attack & defense zones, Running circle model Tactical; Post-hoc Simulation and validation with BWF match statistics 35 [45] Kinect based posture recognition and action evaluation system Badminton athletes and collegiate participants Multi Kinect V2 skeletal and depth data Sensor calibration, Kalman filtering, CNN with attention, Transformer Accuracy up to 98.75%; MAE 0.1364; F1 score 0.6911; 9.27 ms processing time Technical; Real-time and Post-hoc Baseline comparison and ablation study 36 [46] Vision-based badminton stroke recognition and tactical analysis Elite badminton athletes Broadcast and match video footage Improved YOLO-based detection, pose estimation, temporal sequence modelling Stroke classification accuracy above 90%; mAP and F1-score reported Technical and Tactical; Post-hoc Train/Test split, baseline comparison with standard YOLO variants 37 [47] Vision-based badminton posture recognition and training assistance system Badminton players (experimental evaluation cohort) Improved OpenPose (MobileNet backbone) + Particle Filter tracking Improved OpenPose with MobileNet backbone; Particle Filter tracking FPS, Recall, Precision, Accuracy, F1-score, response time, CPU usage Technical; Real-time Experimental evaluation and model comparison 38 [48] Hybrid vision–language framework (ChatMatch) for badminton video understanding Professional players and coach (evaluation); elite match videos Broadcast match videos (20 singles matches) UNet (court segmentation), YOLOv5, ResNet-50 (action recognition), GPT-3.5-based multi-agent system Location accuracy 0.991; Action accuracy 0.902; Gesture accuracy 0.950 Technical and Tactical; Post-hoc Train/Test split; confusion matrices; user-based evaluation. 39 [49] Intelligent badminton handle with multinode MEMS sensors for explainable motion recognition Badminton players (experimental evaluation cohort) Multinode MEMS sensors embedded in racket handle (accelerometer and gyroscope data) CNN-based motion recognition with explainability analysis Recognition accuracy above 90%; precision, recall, F1-score reported Technical; Real-time Experimental validation with controlled stroke trials; comparative evaluation with baseline models 40 [50] Inertial sensor–based badminton swing recognition 10 players; 1,200 swing samples Racket-mounted accelerometer and gyroscope data; VideoBadminton dataset PCA, SVM (grip classification), AdaBoost (six-class swing recognition) Accuracy 95.93%; Precision, Recall, F1-score; AUC up to 0.97 Technical; Post-hoc 5- and 10-fold cross-validation; baseline comparison 41 [51] Grip-force analysis for badminton performance evaluation using flexible pressure sensors 30 male participants (15 beginners, 15 national-level athletes) Flexible piezoresistive pressure sensors embedded in racket handle Signal acquisition and force-feature extraction (Imax, Iave, T); statistical correlation analysis Grip peak force, mean force, duration; performance scores; correlation coefficients (r ≈ −0.6 to −0.7) Technical; Post-hoc Group comparison (beginner vs athlete); correlation analysis; significance testing (p < 0.05) 42 [52] Deep learning and analytics for biomechanics, injury prediction, and tactical modelling Competitive badminton players Match video, wearable sensors, historical data CNN pose estimation; Random Forest (injury). 93% biomechanical accuracy; 90% injury prediction; 15% speed gain; 25% injury reduction Technical and Tactical; Real-time and Post-hoc Controlled evaluation, confusion matrix, tactical performance comparison 43 [53] Motion tracking and decomposition model for badminton training Badminton training participants 3D limb joint coordinate data via motion sensing system 3D decomposition model; Gauss-Newton optimisation Accuracy up to 99.31%; >34% improvement vs LSTM; 32.5% faster response Technical; Real-time Baseline comparison with LSTM; model accuracy and efficiency tests 44 [54] Vision-based badminton stroke detection and performance evaluation Competitive badminton players Match video footage Deep CNN-based detection and temporal sequence modelling Accuracy, Precision, Recall, F1-score reported Technical; Post-hoc Experimental evaluation and baseline comparison 45 [55] Integrated match analysis: detection, tracking, shot classification Elite match videos Broadcast footage; VideoBadminton dataset Modified YOLOv8; court-aware tracking; SlowFast + TimeSformer [email protected] 94.12%; Top-1 77.8%; Top-5 94.24%; 28 FPS Technical and Tactical; Real-time and Post-hoc Train/Test split; baseline comparison; confusion matrix 46 [56] Real-time action recognition and pose estimation with shot refinement Professional match videos (62 matches) Broadcast footage with annotated hits Improved YOLOv8 + 13-keypoint pose model + Kalman filtering Accuracy 90.2%; F1-score 87.5%; Temporal error 2.14 frames Technical and Tactical; Real-time and Post-hoc Baseline comparison; cross-scenario evaluation 47 [57] Stroke recognition using Quantum CNN (QCNN) Competitive players (540 stroke samples) Broadcast videos; OpenPose joint data QCNN vs SVM and 3D-CNN Accuracy and F1 up to 0.965; noise robustness ACC 0.882 Technical; Post-hoc 70/15/15 split; baseline comparison; noise evaluation 48 [58] Deep learning–based badminton action recognition with multimodal fusion and computational optimisation 19 university-level athletes (VideoBadminton dataset; 7,822 clips) VideoBadminton dataset (annotated match videos) 3D-CNN, Bi-LSTM, Video Swin Transformer; multimodal fusion; transfer learning Accuracy up to 0.967; F1 up to 0.953; inference time 10.8–13.1 ms; reduced FLOPs and memory usage Technical; Real-time and Post-hoc 70/15/15 split; comparison with temporal graph network and other baseline models; efficiency and performance benchmarking 49 [59] Badminton action recognition with multimodal fusion and computational optimisation 19 university athletes (VideoBadminton dataset; 7,822 clips; 18 action classes) Annotated match videos (VideoBadminton) 3D-CNN, Bi-LSTM, Video Swin Transformer; multimodal fusion; transfer learning Accuracy up to 0.967; F1 up to 0.953; MSE 0.024–0.037; inference 10.8–13.1 ms Technical; Real-time and Post-hoc 70/15/15 split; comparison with temporal graph convolutional network and other multimodal networks; efficiency benchmarking 50 [60] Key frame detection in badminton swings for educational feedback University students (102 videos; 143 in intervention) Tablet video; MediaPipe 3D skeleton data ST-GCN; MLP; temporal correction algorithm Recall; F1-score; Average Positional Error Technical; Real-time and Post-hoc Cross-validation; intervention study; statistical evaluation 51 [3] Tactical badminton analysis using computer vision and Retrieval-Augmented Generation (RAG)-anhanced large language model (LLM) Elite singles matches (>750 rallies) 1080p broadcast videos (BWF) YOLOv8; MediaPipe pose; multi-cue shot detection; RAG-LLM (LangChain + FAISS) mAP 0.94 (player); 0.83 (shuttle); F1 0.92 Tactical; Post-hoc Quantitative evaluation; statistical testing; expert validation Table II shows the range of research designs in AI-based badminton player assessment. Most studies employ controlled experiments to analyse technical execution, biomechanics, and tactics, though often with reduced ecological validity. Observational approaches using broadcast or archived footage capture authentic match play but are limited by video quality. A smaller number of works focus on algorithmic innovation or feasibility studies addressing deployment challenges. The table also shows a concentration on elite and collegiate athletes, with fewer studies on recreational or youth groups, such as Wang et al. [9] for elite tactical analysis, Sinadia and Murwantara [6] for collegiate profiling, Liu and Liang [11] for recreational cohorts, and Amudhan et al. [41] for youth shuttlecock detection. This emphasis provides insights for high-performance contexts but limits generalisability to the wider badminton community, which is dominated by recreational and developing players [28,41,43]. Most AI-based badminton player assessment studies rely on video footage as the primary data source [3,5,7–9], reflecting its accessibility through broadcasts and archives and its minimal burden on players. A smaller set employs wearable sensors to capture biomechanical or physiological data [51], with only a few combining video and sensors to leverage spatial–temporal resolution alongside on-body kinematics. For instance, Wang et al. [9] introduced the ShuttleSet dataset with annotated video for tactical analysis, Van Herbruggen et al. [13] integrated inertial and ultra-wideband sensors with video to study match strategies, and Zheng and Chen [32] fused video, sensors, and historical records to support both tactical assessment and injury prevention. While video remains the most practical modality, hybrid approaches offer promising opportunities to enhance accuracy and contextual richness in future badminton analytics. 3.1.3. AI Methodologies for Badminton Player Assessment AI research in badminton can be categorised into four principal application areas: stroke analysis, movement-pattern analysis, spatio-temporal or tactical analysis, and multi-modal technical evaluation. The literature is predominantly concentrated on stroke-level modelling, where computer vision, deep learning, and pose estimation techniques are applied to detect and classify strokes from video data. Movement-pattern studies focus on footwork recognition, trajectory tracking, and positional dynamics to characterise court coverage and recovery behaviour. Spatio-temporal or tactical approaches examine rally sequences, shot influence, and strategic decision-making processes, extending analysis beyond isolated actions to match-level contexts. Multi-modal technical evaluation integrates heterogeneous data sources, including video, wearable sensors, and biomechanical signals, to provide more comprehensive performance assessment. Overall, current research emphasises visual stroke detection, while integrated and strategy-oriented evaluation frameworks remain comparatively less developed. Stroke analysis tasks in badminton frequently adopt deep learning classifiers such as 2D- CNNs, 3D-CNNs, and transformer-based sequence models to recognise stroke categories including smashes, drops, clears, and drives [5,18,30,39]. CNN-based approaches excel at capturing spatial features from individual frames, while 3D-CNNs extend this to spatio-temporal volumes, improving recognition of subtle motion cues. Transformer architectures and temporal CNNs further enhance performance by modelling long-range dependencies across rally sequences, making them effective in distinguishing strokes with similar visual appearance but different execution timing. Reported accuracies often exceed 85–90%, highlighting the maturity of these models for technical skill classification [10,13,17,18,23,29]. Movement-pattern analysis, on the other hand, builds on pose estimation outputs from frameworks such as OpenPose and MediaPipe, which generate joint keypoints used as input features for downstream models [10–12,31]. Recurrent neural networks, long short-term memory (LSTM), and graph convolutional networks are commonly employed to capture temporal dependencies and relational dynamics between body joints, enabling assessment of agility, footwork recovery, and tactical positioning [22]. Some studies incorporate You Only Look Once (YOLO)-based detectors for robust player localisation before pose extraction, ensuring consistent tracking under broadcast video conditions [9,12,18,28,31]. Movement-pattern and rally-sequencing analyses in badminton performance span spatial, temporal, and cognitive domains, including court coverage [22,48], footwork sequencing [12,23], recovery positioning [10], and contextual rally dynamics [7–9]. Spatial mapping and trajectory overlays produce heat-map visualisations aligned with expert evaluations [12,23], while pose-estimation models reliably classify six-corner footwork patterns [9]. Recovery-position models based on skeletal features achieve high accuracy but vary in predictive stability, reflecting the situational complexity of tactical decision-making [10]. Rally-sequencing studies highlight strategic shot-to-shot transitions that distinguish playing styles and outcomes [7–9], with machine learning approaches including decision tree (DT), random forest (RF), and graph models used to evaluate shot selection in rally contexts [7,22]. Integrating perceptual–cognitive markers, Tan and Teoh [27] combined Quiet Eye metrics with biomechanical features in a neural network, achieving over 85% accuracy in predicting shot outcomes and identifying gaze duration and onset timing as critical predictors. 3.1.4. Assessment Validation Approaches AI-based badminton player assessment employs four main categories of performance indicators: biomechanical, shot quality, tactical, and movement-related metrics. Biomechanical measures, such as joint angles and segment velocities, are validated against expert annotations and biomechanical reference models, and these features are subsequently encoded as kinematic vectors for classifier training [10,12,40]. Shot-quality indicators, including accuracy, consistency, and power, are benchmarked using supervised classification pipelines with cross-validation, and they function as target variables in stroke recognition models [26]. Tactical metrics, such as shot selection and court positioning, are assessed through predictive accuracy and correlations with rally outcomes [55], and they are typically modelled as sequential dependencies within recurrent or transformer-based architectures [25]. Movement-related measures, notably footwork efficiency and court coverage, are validated using pose-tracking precision and trajectory error, and they are structured into graph or temporal representations for learning locomotor strategies [12]. Within AI-based badminton assessment, existing frameworks vary according to their analytical focus and temporal orientation. Most are designed as technical post-analysis systems, emphasising stroke mechanics and biomechanical precision once matches have concluded [7,17]. A smaller body of work addresses real-time technical assessment, where immediate feedback is achieved under favourable computational conditions [40]. Tactical strategy frameworks are reported less frequently and are generally retrospective, analysing rally sequences and decision-making behaviour through post-match data [8]. Real-time tactical guidance systems are rare, though studies demonstrate their capacity to provide adaptive decision support during active play [9,12,23,38]. Hybrid frameworks that combine technical and tactical elements across flexible timelines are limited but illustrate attempts to capture mechanical precision alongside contextual dynamics [24,28]. 3.2. Common Challenges in AI-Based Badminton Assessment Research (RQ 2) A recurrent theme across the reviewed literature is the persistence of technical challenges that constrain the reliability and scalability of AI-based badminton assessment systems. Data acquisition difficulties are the most frequently reported, particularly in relation to shuttlecock tracking and court reference extraction. Hsu et al. [40] highlighted the limitations of monocular video, where occlusion and high shuttle speed hinder consistent localisation, thereby impairing downstream analytical processes. Similarly, Wei and Weng [37] demonstrated that court line extraction from broadcast footage remains sensitive to lighting variation and camera zoom, reducing the stability of detection pipelines. These examples underscore how the quality and consistency of raw input data continue to represent a major obstacle for performance assessment frameworks. Integration complexity emerges as another common limitation, particularly in systems that combine multiple data modalities. Ghosh et al. [28] emphasised the engineering burden of synchronising heterogeneous streams within the DeCoach framework, while Van Herbruggen et al. [13] noted similar challenges when aligning inertial sensor data with high-frame-rate video to maintain biomechanical fidelity. Closely related are concerns of generalisability, as many models perform well within narrowly defined datasets but degrade when applied to different playing styles or populations. Baclig et al. [34] demonstrated this difficulty when multi-player tracking models developed in squash contexts lost accuracy in badminton rallies, and Zhi et al. [39] reported analogous domain-shift effects in stroke recognition tasks. Further limitations involve computational efficiency, real-time processing, and tracking precision. Yang et al. [18] explored model pruning and mixed-precision computation to preserve analytic accuracy while remaining within hardware constraints, whereas Salim et al. [26] designed adaptive inference pipelines that modulate processing load according to rally tempo. Tracking accuracy has also been identified as a recurring difficulty, as even minor errors in court-line detection can propagate across extended rallies, producing cumulative distortions in positional analysis [38,39]. These technical challenges illustrate the ongoing tension between methodological sophistication, practical feasibility, and ecological validity in AI-based badminton performance assessment. 3.3 Interpretability and Usability of AI Outputs in Badminton Player Assessment (RQ3) 3.3.1 Model Interpretability and Explainable Vision–Language Frameworks Only 4 of the 51 reviewed studies incorporated explicit interpretability mechanisms. Wang [8] provided frame-level predictions with Grad-CAM heat maps that highlight the court regions influencing each decision, whereas Chen et al. [9] attached SHAP value attributions to rally-phase classifications. These visual or feature-based explanations transform otherwise opaque predictions into reasoned assessments that align with coaching terminology. Despite their demonstrated value, the limited uptake indicates a substantial gap between algorithmic precision and user confidence. More recently, interpretability has also been extended beyond feature attribution towards structured vision–language reasoning. The ChatMatch framework [48] integrates meta-feature extraction, rule-based knowledge decoding, and LLM agents to generate structured and unstructured explanations of match events. By transforming spatial, action, and gesture recognition outputs into descriptive rally narratives and statistical summaries, the system provides traceable reasoning pathways between visual evidence and analytical conclusions [3]. This hybrid architecture illustrates a shift from post hoc visualisation towards interactive, explainable inference mechanisms that support professional query-based analysis. 3.3.2 Implications for Coaching Practice Evidence from practice-focused evaluations shows that interpretable outputs improve both usability and acceptance. DeCoach [28] is used to quantify vertical growth stages objectively. Wang et al. [29] integrated bounding-box projections from their YOLO-HGNet detector with kinematic graphs, allowing for inspection of fast exchanges that are difficult to analyse visually. At larger scales, the ShuttleSet framework introduced by Wang et al. [22] processed more than thirty-six thousand strokes and provides dashboard summaries that shorten the routine assessment of approximately five hundred players. In a complementary line of work, Zhang and Zhong [45] developed a Kinect-based auxiliary training system that integrates multi-sensor skeletal fusion, gesture recognition, and hierarchical action evaluation to support structured badminton teaching. Their system demonstrated high motion-recognition accuracy and enabled real-time corrective feedback, illustrating how sensor-driven pose analysis can be embedded into pedagogical workflows. 3.3.3 Applications for Competition Organisation Explainable assessment outputs also facilitate fairer competition structures. Sinadia and Murwantara [6] demonstrated that machine-learning profiles linked to positional heat maps can refine club assessment systems, while Asriani et al. [17] proved that spatio-temporal rally analysis is more sensitive than fixed-period reviews when reclassifying players. Xie et al. [44] further reported that pose-aware analytics enable organisers to design tournaments tailored to diverse tactical styles. Collectively, these studies suggest that transparent analytics support equitable seeding, responsive regrading, and data-informed scheduling. 3.3.4 Player Development Implications For athletes, interpretable feedback promotes self-directed improvement. A study by Liu and Liang [11] presents an action-recognition tool that players use between formal sessions to monitor technique independently. Yuan et al. [7] proposed a multi-feature framework that blends technical and tactical indices to personalise training priorities. Tan and Teoh [27] combined neural network shot predictions with accurate metrics, providing objective evidence of skill progression that can enhance motivation. These findings indicate that accessible explanations can extend the impact of AI systems beyond coaching sessions, supporting continuous learning and goal-oriented practice. 3.3.5 Methodological Effectiveness Many of the studies reviewed highlight recurring themes in performance testing that are increasingly relevant to practical coaching. One area with strong potential is the use of pose estimation to analyse player movements. For example, the enhanced YOLO model developed by Yang et al. [18] improved the accuracy of posture detection in badminton players by more than eight percent compared to earlier systems. This improvement was achieved by adjusting how the model processes visual information, resulting in greater efficiency and reliability. However, the availability of badminton-specific datasets remains limited, particularly for multi-view capture and synchronised multi-modal data [23], and addressing these constraints is essential for achieving more robust and generalisable pose-estimation models. In addition to technical evaluation, some studies have explored the use of spatial and temporal features to support player assessment, including measurements of body position and timing. These models offer promising frameworks for translating movement data into meaningful performance insights. However, evaluating tactical decision-making remains more challenging. As shown in the ShuttleSet study by Wang [22], analysing strategic behaviour requires substantial human annotation and expert interpretation. This makes tactical assessment harder to scale and apply consistently. These limitations point to the need for stronger collaboration between AI developers and coaching practitioners to create systems that are not only accurate but also practical and intuitive for use in real sport settings. 4. Discussion This review set out to clarify three questions: which AI techniques work best for badminton player assessment, what methodological hurdles still stand in the way, and how clearly current systems communicate their findings. Thirty-four primary studies, summarised in Table II, show that computer vision pipelines now dominate the field. Yet, they sit alongside movement-pattern analytics, spatio-temporal sequence models, and a small but important group of multi-modal systems. Together, these works sketch a discipline that has matured rapidly but remains uneven in scope and depth. Modern vision models that track body landmarks and shuttle flight now achieve remarkable accuracy, with Yang et al. [ 18 ] reporting stroke-recognition rates above 95%, demonstrating rapid progress in automated analysis. By combining image cues with temporal contexts, these systems can assess not only what players do but also when and why actions matter within a rally, providing coaches with quantitative insights into technique, tactical rhythm, and fatigue. Yet, despite matching or exceeding human reliability in routine evaluations, many studies acknowledge that performance often degrades under varying recording conditions, highlighting the gap between technical promise and practical robustness and raising the critical question of what barriers limit broader adoption. Data collection emerges as the most frequently reported obstacle in AI-based badminton player assessment. Hsu et al. [ 40 ] observed that the shuttlecock often disappears from a single camera’s field of view due to its high velocity, disrupting otherwise reliable analytic pipelines. Integration complexity represents another recurrent challenge, with Ghosh et al. [ 28 ] documenting the substantial effort needed to synchronise video streams, inertial sensors, and match statistics within the DeCoach framework. Additional constraints arise from limited generalisability, computational overhead, and real-time latency, all of which highlight the contrast between controlled laboratory benchmarks and the variable conditions of competition environments characterised by fluctuating lighting, multiple camera angles, and heterogeneous player styles. Only two studies, Wang [ 8 ] and Chen et al. [ 9 ], integrated explanation tools such as Grad-CAM or SHAP, showing that even simple visual cues boost coach confidence and player engagement. More recent systems, including ChatMatch [ 48 ] and Court-to-Conversation [ 3 ], integrated RAG and LLM components to transform detected visual events into structured, query-driven analytical feedback. By grounding language outputs in retrieved match data and recognised actions, these frameworks establish traceable links between raw visual evidence and narrative explanations, thereby supporting coach-player interaction. Future progress depends on open multi-modal datasets, models with domain adaptation and uncertainty estimation, and interpretability as a standard rather than a novelty. Longitudinal studies that track how AI-feedback shapes skill development and performance will then provide the evidence needed for widespread adoption. 5. Conclusion This systematic review examined the current development of AI applications in badminton player assessment, highlighting both methodological diversity and several key limitations. Although AI systems offer the potential for objective, consistent, and multi-dimensional performance evaluation, their adoption remains limited due to weak theoretical grounding, insufficient methodological rigour, and limited contextual adaptability. Existing studies employ various approaches, including computer vision for stroke recognition, movement analysis, spatio-temporal modelling, and multi-modal frameworks. Future research should strengthen the theoretical foundation by linking computational outputs with established models of skill acquisition and athlete development, while also developing standardised performance metrics to enable meaningful comparisons across studies. A promising future trajectory lies in developing interactive and explainable AI systems supported by LLMs, enabling performance analysis to be translated into natural, context-aware explanations that coaches and players can easily interpret and utilise. Declarations Competing Interests The authors declare that they have no known competing financial or non-financial interests that could have appeared to influence the work reported in this paper. Funding The authors received no specific funding for this work. Author Contribution R.C. and F.U.D. designed the study and supervised a student group in carrying out the initial analysis. C.A. carried out further analysis and wrote the manuscript. All authors reviewed, edited and approved the final manuscript. Acknowledgement The authors would like to acknowledge the initial work carried out by a student group led by Utsab Gyawali, which contributed to the early exploration of this research topic. References Wörner EA, Safran MR. Racquet Sports: Tennis, Badminton, Racquetball, Squash. Specific Sports-Related Injuries, Cham: Springer International Publishing; 2021, p. 431–46. https://doi.org/10.1007/978-3-030-66321-6_30. Woo T-MT, Alam F. Comparative aerodynamics of synthetic badminton shuttlecocks. Sports Engineering 2018;21:21–9. https://doi.org/10.1007/s12283-017-0241-2. Bharadwaj K, Srinivasa G. Court to conversation: Tactical badminton analysis via computer vision and RAG-enhanced LLMs. Knowl Based Syst 2026;333:115027. https://doi.org/10.1016/j.knosys.2025.115027. Jekauc D, Burkart D, Fritsch J, Hesenius M, Meyer O, Sarfraz S, et al. Recognizing affective states from the expressive behavior of tennis players using convolutional neural networks. Knowl Based Syst 2024;295:111856. https://doi.org/10.1016/j.knosys.2024.111856. Ghosh A, Singh S, Jawahar C V. Towards structured analysis of broadcast badminton videos. Proceedings of the IEEE Winter Conference on Applications of Computer Vision 2018:296–304. https://doi.org/10.1109/WACV.2018.00039. Sinadia HE, Murwantara IM. Sports analytics: A comparison of machine learning performance for profiling badminton athlete. Proceedings of the International Conference on Technology Innovation and Its Applications 2022. https://doi.org/10.1109/ICTIIA54654.2022.9935852. Yuan H, Wang Y, Yang K, Bin Y. Prediction model and technical and tactical decision analysis of women’s badminton singles based on machine learning. PLoS One 2024;19. https://doi.org/10.1371/journal.pone.0312801. Wang KD. Enhancing badminton player performance via a closed‑loop AI approach: Imitation, simulation, optimization, and execution. Proceedings of the International Conference on Information and Knowledge Management 2023:5189–92. https://doi.org/10.1145/3583780.3616001. Chen WT, Wu HY, Shih YA, Wang CC, Wang YS. Exploration of player behaviours from broadcast badminton videos. Computer Graphics Forum 2023;42:267–79. https://doi.org/10.1111/cgf.14786. Krishnaram N, Ahamed J, Sathyamoorthy N, Sandaruwan KD, Athapaththu AMKB. Skeletal point analysis to determine the accuracy of forehand smash shots played by badminton players. J Natl Sci Found 2024;52:125–42. https://doi.org/10.4038/jnsfsr.v52i1.12141. Liu J, Liang B. An action recognition technology for badminton players using deep learning. Mobile Information Systems 2022. https://doi.org/10.1155/2022/3413584. Jannet K, Sunar MS, Molla MMI, Bin As’Ari MA. A deep learning approach to badminton player footwork detection based on YOLO models: A comparative study. Proceedings of the IEEE International Conference on Signal and Image Processing Applications 2024. https://doi.org/10.1109/ICSIPA62061.2024.10686537. Van Herbruggen B, Fontaine J, Simoen J, De Mey L, Peralta D, Shahid A, et al. Strategy analysis of badminton players using deep learning from IMU and UWB wearables. Internet of Things 2024;27. https://doi.org/10.1016/j.iot.2024.101260. Marques A, Gouveira ÉR, Peralta M, Martins J, Venturini J, Henriques-Neto D, et al. Cardiorespiratory fitness and telomere length: a systematic review. J Sports Sci 2020;38:1690–7. https://doi.org/10.1080/02640414.2020.1754739. Zhou D, Keogh JWL, Ma Y, Tong RKY, Khan AR, Jennings NR. Artificial intelligence in sport: A narrative review of applications, challenges and future trends. J Sports Sci 2025:1–16. https://doi.org/10.1080/02640414.2025.2518694. Vickery-Howe DM, Bonanno DR, Dascombe BJ, Drain JR, Clarke AC, Hoolihan B, et al. Physiological, perceptual, and biomechanical differences between treadmill and overground walking in healthy adults: A systematic review and meta-analysis. J Sports Sci 2023;41:2088–120. https://doi.org/10.1080/02640414.2024.2312481. Asriani F, Azhari A, Wahyono W. Improving badminton action recognition using spatio‑temporal analysis and a weighted ensemble learning model. Computers, Materials and Continua 2024;81:3079–96. https://doi.org/10.32604/cmc.2024.058193. Yang W, Jiang M, Fang X, Shi X, Guo Y, Al‑qaness MAA. A high‑precision and efficient method for badminton action detection in sports using You Only Look Once with Hourglass Network. Eng Appl Artif Intell 2024;137:109177. https://doi.org/10.1016/j.engappai.2024.109177. Cao Z, Liao T, Song W, Chen Z, Li C. Detecting the shuttlecock for a badminton robot: A YOLO based approach. Expert Syst Appl 2021;164:113833. https://doi.org/10.1016/j.eswa.2020.113833. Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021:n71. https://doi.org/10.1136/bmj.n71. Helbach J, Hoffmann F, Pieper D, Allers K. Reporting according to the preferred reporting items for systematic reviews and meta-analyses for abstracts (PRISMA-A) depends on abstract length. J Clin Epidemiol 2023;154:167–77. https://doi.org/10.1016/j.jclinepi.2022.12.019. Wang WY, Huang YC, Ik TU, Peng WC. ShuttleSet: A human‑annotated stroke‑level singles dataset for badminton tactical analysis. Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining 2023:5126–36. https://doi.org/10.1145/3580305.3599906. Luo J, Hu Y, Davids K, Zhang D, Gouin C, Li X, et al. Vision‑based movement recognition reveals badminton player footwork using deep learning and binocular positioning. Heliyon 2022;8. https://doi.org/10.1016/j.heliyon.2022.e10089. Ding N, Takeda K, Fujii K. Deep reinforcement learning in a racket sport for player evaluation with technical and tactical contexts. IEEE Access 2022;10:54764–72. https://doi.org/10.1109/ACCESS.2022.3175314. Ashfaq F, Jhanjhi NZ, Khan NA. Badminton player’s shot prediction using deep learning. Lecture Notes in Bioengineering 2023:233–43. https://doi.org/10.1007/978-981-99-0297-2_19. Salim MA, Williamdy W, Steven E. AI‑powered badminton shot classification. Proceedings of the International Conference on Technology Innovation and Its Applications 2024. https://doi.org/10.1109/ICTIIA61827.2024.10761547. Tan S, Teoh TT. Predicting shot accuracy in badminton using quiet eye metrics and neural networks. Applied Sciences 2024;14:9906. https://doi.org/10.3390/app14219906. Ghosh I, Ramasamy Ramamurthy S, Chakma A, Roy N. DeCoach: Deep learning‑based coaching for badminton player assessment. Pervasive Mob Comput 2022;83. https://doi.org/10.1016/j.pmcj.2022.101608. Wang WY, Chan TF, Peng WC, Yang HK, Wang CC, Fan YC. How is the stroke? Inferring shot influence in badminton matches via long short‑term dependencies. ACM Trans Intell Syst Technol 2022;14. https://doi.org/10.1145/3551391. Peng W, Zheng W. An AI-based badminton smash detection using residual-shuffle network optimized based on upgraded pufferfish optimizer. Ain Shams Engineering Journal 2025;16:103414. https://doi.org/10.1016/j.asej.2025.103414. Wu Y, Chen Z, Zhang H, Yang Y, Yi W. Enhanced Pose Estimation for Badminton Players via Improved YOLOv8-Pose with Efficient Local Attention. Sensors 2025;25:4446. https://doi.org/10.3390/s25144446. Zheng Y, Chen E. Harnessing deep learning and advanced analytics to revolutionise badminton performance. International Journal of Information and Communication Technology 2025;26. https://doi.org/10.1504/IJICT.2025.10070826. Rahmad NA, Sufri NAJ, Muzamil NH, As’ari MA. Badminton player detection using faster region convolutional neural network. Indonesian Journal of Electrical Engineering and Computer Science 2019;14:1330–5. https://doi.org/10.11591/ijeecs.v14.i3.pp1330-1335. Baclig MM, Ergezinger N, Mei Q, Gul M, Adeeb S, Westover L. A deep learning and computer vision based multi‑player tracker for squash. Applied Sciences 2020;10. https://doi.org/10.3390/app10248793. Liu P, Wang J-H. MonoTrack: Shuttle trajectory reconstruction from monocular badminton video. Proceedings of the IEEE Computer Vision and Pattern Recognition Workshops 2022:3512–21. https://doi.org/10.1109/CVPRW56347.2022.00395. Wu F, Wang Q, Bian J, Ding N, Lu F, Cheng J, et al. A survey on video action recognition in sports: Datasets, methods and applications. IEEE Trans Multimedia 2023;25:7943–66. https://doi.org/10.1109/TMM.2022.3232034. Wei CT, Weng SK. A court line extraction algorithm for badminton tournament videos with horizontal line projection learning. IET Image Process 2023;17:2907–24. https://doi.org/10.1049/ipr2.12838. Zhang L, Dai H. Motion trajectory tracking of athletes with improved depth information‑based KCF tracking method. Multimed Tools Appl 2023;82:26481–93. https://doi.org/10.1007/s11042-023-14929-6. Zhi J, Sun Z, Zhang R, Zhao Z. Badminton video action recognition based on time network. Journal of Computational Methods in Sciences and Engineering 2023;23:2739–52. https://doi.org/10.3233/JCM-226889. Hsu YH, Yu CC, Cheng HY. Enhancing badminton game analysis: An approach to shot refinement via a fusion of shuttlecock tracking and hit detection from monocular camera. Sensors 2024;24. https://doi.org/10.3390/s24134372. Amudhan AN, Vrajesh SR, Sudheer AP, Lijiya A. RFSOD: A lightweight single‑stage detector for real‑time embedded applications to detect small‑size objects. J Real Time Image Process 2022;19:133–46. https://doi.org/10.1007/s11554-021-01170-3. Tan YE, See J, Abdullah J, Wang LK, Ban KW. A deep learning based framework for badminton rally outcome prediction. Proceedings of the International Symposium on Intelligent Signal Processing and Communication Systems 2022. https://doi.org/10.1109/ISPACS57703.2022.10082764. Gonzalez MR, Sandoval JC, Tostado GF, Santamaria SS, San Blas HS, Gonzalez G V, et al. Development of a stroke detection system for racket sports. Proceedings of the International Conference on Disruptive Technologies, Tech Ethics, and Artificial Intelligence 2023:320–8. https://doi.org/10.1007/978-3-031-38344-1_31. Xie Y, Xie X, Xia H, Zhao Z. Algorithmic study on position and movement method of badminton doubles. Sci Program 2022. https://doi.org/10.1155/2022/6422442. Zhang L, Zhong P. Feather teaching and training based on Kinect sensor and gesture recognition technology. Systems and Soft Computing 2025;7:200349. https://doi.org/10.1016/j.sasc.2025.200349. Peng W, Zheng W. An AI-based badminton smash detection using residual-shuffle network optimized based on upgraded pufferfish optimizer. Ain Shams Engineering Journal 2025;16:103414. https://doi.org/10.1016/j.asej.2025.103414. Jiang K. Optimization study of badminton sports training system based on MoileNet OpenPose lightweight human posture estimation model. Entertain Comput 2025;54:100975. https://doi.org/10.1016/j.entcom.2025.100975. Zhang J, Han D, Han S, Li H, Lam W-K, Zhang M. ChatMatch: Exploring the potential of hybrid vision–language deep learning approach for the intelligent analysis and inference of racket sports. Comput Speech Lang 2025;89:101694. https://doi.org/10.1016/j.csl.2024.101694. Li J, Fan Y, Chen R, Liang S, Feng Y, He Y, et al. An Intelligent Badminton Handle With Multinode MEMS Sensors for Explainable Motion Recognition. IEEE Internet Things J 2025;12:37022–34. https://doi.org/10.1109/JIOT.2025.3580545. Shih C-Y, Lin Y-T, Chen W, Huang J-C. SVM-Adaboost based badminton offensive movement parsing technique. Signal Image Video Process 2025;19:280. https://doi.org/10.1007/s11760-025-03865-7. Liu D, Che L, Qi F, Zhang K, Cao D. Evaluating the sports performance of badminton players based on grip strength of the real hitting scenario. Sci Rep 2026;16:5055. https://doi.org/10.1038/s41598-026-35596-1. Chen E, Zheng Y. Harnessing deep learning and advanced analytics to revolutionise badminton performance. International Journal of Information and Communication Technology 2025;26:61–79. https://doi.org/10.1504/IJICT.2025.146097. Liu M. Badminton Technical Training Analysis System Based on Intelligent Motion Tracking Decomposition Model. 2025 International Conference on Algorithm, Artificial Intelligence and Computer Vision (AAICV), IEEE; 2025, p. 01–4. https://doi.org/10.1109/AAICV66571.2025.00013. Tan and J. Komar, “How Data Can Help You Beat Your Next Opponent: Applications of Artificial Intelligence for Performance Analysis in Badminton,” in Sports Data Analytics: Techniques, Applications, and Innovations, M. S. Raval, T. Kaya, N. S. Artan, and C. Taber, Eds. Singapore: Springer Nature Singapore, 2026, pp. 141–162. https://doi.org/10.1007/978-981-95-5132-3_8. Singha KAK, Ali MA, Acharjee T, Debnath P, Ahmadullah S. Badminton Match Analysis Along With Players and Shuttlecock Setection Using Hybrid CNN. 2025 IEEE Silchar Subsection Conference (SILCON), IEEE; 2025, p. 1–6. https://doi.org/10.1109/SILCON67893.2025.11327275. Luo L, Zhang L, Liu H. An integrated deep learning framework for real-time badminton action recognition and pose estimation with shot-level semantic refinement. Australian Journal of Electrical and Electronics Engineering 2025:1–20. https://doi.org/10.1080/1448837X.2025.2553948. Zhu X, Liu L, Huang J, Chen G, Ling X, Chen Y. The analysis of motion recognition model for badminton player movements using machine learning. Sci Rep 2025;15:19030. https://doi.org/10.1038/s41598-025-02771-9. Ma H, Zhang F, Liang N. Development and Training Strategy of Badminton Action Recognition System Under the Background of Artificial Intelligence. IEEE Access 2025;13:71577–85. https://doi.org/10.1109/ACCESS.2025.3563302. Chang Y. A method of badminton video motion recognition based on adaptive enhanced AdaBoost algorithm. Int J Biom 2025;17:86–101. https://doi.org/10.1504/IJBM.2025.143715. Hsu J-H, Lee C-C, Chang J-Y, Lee D-S. Key Frame Detection in Badminton Swings and Its Application to Physical Education. IEEE Access 2025;13:91248–62. https://doi.org/10.1109/ACCESS.2025.3572105. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9105158","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":607427203,"identity":"05df9293-9d72-401e-987b-3b75b3149aa0","order_by":0,"name":"Christian Arthur","email":"","orcid":"","institution":"The University of New England","correspondingAuthor":false,"prefix":"","firstName":"Christian","middleName":"","lastName":"Arthur","suffix":""},{"id":607427204,"identity":"a7438293-ac47-4411-94da-a9574bdabcdc","order_by":1,"name":"Raymond Chiong","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyElEQVRIiWNgGAWjYFACHoYDQFIOwmEjQYsxaVpAILGBaC3m7WcPHvi4ozZ9w+3mAwwfyg4z8M9IwK9F5kxewsGZZ47nbrhzLIFxxrnDDBI3CGiRkOAxOMzbdix3w40cA2betsMMDMRqSTe4kf+B+S9QizyRWmoSDG7kMDAzArUYENTCk2NwcGbbAcOZN9IMDvacS+cxPPOAgBb2M8YfPrbVyfPdSH744EeZtZzccQK2QMFhMHmAARZNRIA6YhWOglEwCkbBSAQAcFVHa1mI5MYAAAAASUVORK5CYII=","orcid":"","institution":"The University of New England","correspondingAuthor":true,"prefix":"","firstName":"Raymond","middleName":"","lastName":"Chiong","suffix":""},{"id":607427205,"identity":"5c790f06-8cde-40a7-a203-5b616054a97a","order_by":2,"name":"Fareed Ud Din","email":"","orcid":"","institution":"The University of New England","correspondingAuthor":false,"prefix":"","firstName":"Fareed","middleName":"Ud","lastName":"Din","suffix":""},{"id":607427206,"identity":"1a7b6c04-ffdf-4cec-8088-469c5bf09f94","order_by":3,"name":"Farshid Hajati","email":"","orcid":"","institution":"The University of New England","correspondingAuthor":false,"prefix":"","firstName":"Farshid","middleName":"","lastName":"Hajati","suffix":""},{"id":607427207,"identity":"122d47f7-e813-4f2e-ae26-a30a27fb28db","order_by":4,"name":"Subrata Chakraborty","email":"","orcid":"","institution":"The University of New England","correspondingAuthor":false,"prefix":"","firstName":"Subrata","middleName":"","lastName":"Chakraborty","suffix":""},{"id":607427208,"identity":"63005009-da1c-4975-8c6d-a7bd80998995","order_by":5,"name":"David Paul","email":"","orcid":"","institution":"The University of New England","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"","lastName":"Paul","suffix":""},{"id":607427209,"identity":"196cd8f8-2269-4b0b-80a4-ac7bafc4ac4e","order_by":6,"name":"Mitchell Welch","email":"","orcid":"","institution":"The University of New England","correspondingAuthor":false,"prefix":"","firstName":"Mitchell","middleName":"","lastName":"Welch","suffix":""},{"id":607427210,"identity":"308b654d-8bee-4e5b-9ab5-1c3b2044a2ff","order_by":7,"name":"Robert G. Crowther","email":"","orcid":"","institution":"The University of New England","correspondingAuthor":false,"prefix":"","firstName":"Robert","middleName":"G.","lastName":"Crowther","suffix":""}],"badges":[],"createdAt":"2026-03-12 13:08:43","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9105158/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9105158/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104877171,"identity":"2f210021-9b7a-4d5a-bfd8-b1b1273e259b","added_by":"auto","created_at":"2026-03-18 08:46:30","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":193432,"visible":true,"origin":"","legend":"\u003cp\u003ePRISMA flow diagram illustrating the systematic study selection process.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9105158/v1/9c5a8cd062c9ba9c9adf103b.png"},{"id":104877173,"identity":"c64eaf71-8ac5-4c77-970f-8268f6041d10","added_by":"auto","created_at":"2026-03-18 08:46:30","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":195758,"visible":true,"origin":"","legend":"\u003cp\u003eAnnual publication trend of AI applications in badminton analytics from 2015 to 2025.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9105158/v1/8e545e0f0bd2e90c3e833692.png"},{"id":104877172,"identity":"cd0a3aa1-c8d3-4e94-957c-89db564763f4","added_by":"auto","created_at":"2026-03-18 08:46:30","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":48731,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of methodological categories in AI-based badminton assessment studies.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9105158/v1/a5cb1177ec429e0163ef1a6a.png"},{"id":105562688,"identity":"9c1f2bac-2317-418d-9533-4fa625e341e1","added_by":"auto","created_at":"2026-03-27 12:44:05","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1258380,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9105158/v1/e46c1772-e796-43ad-944f-9bf67662f234.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Making the World’s Fastest Racket Sport even Better: A Systematic Review of Artificial Intelligence-based Objective Player Performance Assessment in Badminton","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eSince its inclusion in the Olympic Games in 1992, badminton has developed into a highly competitive racquet sport with a well-established professional structure [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. The sport is widely recognised as the fastest racquet sport, characterised by extremely rapid rallies and shuttle speeds that demand quick reactions, precise technical execution, and continuous tactical adaptation. It is played recreationally by an estimated 200\u0026nbsp;million people worldwide, with more than 7,000 athletes competing annually in hundreds of national and international events [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTraditionally, coaches, selection committees, and technical administrators have based badminton player assessment on subjective evaluation [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. While capitalising on the domain expertise of the practitioner, this approach can lead to inconsistent and biased approaches to player development, competitive structuring, and participant satisfaction [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. The subjective nature of traditional assessment models has spurred increasing interest in more objective, data-driven approaches to player evaluation [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCompared to other sports, badminton is a complex one to assess because of its multiple technical requirements, fast pace of play, and the presence of multiple skill components [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. To be proficient in multiple domains, players must have demonstrated proficiency in stroke techniques, court movement, tactical decision-making, and competitive temperament [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. However, these skills are integrated at speeds much faster than other racket sports [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn most countries worldwide, club-level badminton associations use classification systems to divide players into different skill categories to help establish appropriate competitive structures and training organisation [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Despite this, the subjectivity of assessment often undermines the effectiveness of these systems, resulting in inconsistent assessment, player disputes, and the ability to manipulate for competitive advantage [\u003cspan additionalcitationids=\"CR15\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe integration of artificial intelligence (AI) technologies provides promising pathways to address these challenges by providing a systematic, data-driven multi-factor analysis of player performance [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. With recent advances in computer vision, machine learning, and data analytics, opportunities exist for more objective assessment frameworks [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. However, such systems are theoretically incomplete, methodologically unexplored, and empirically untested [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis review studies AI approaches for badminton player performance assessment. It evaluates their reported accuracy and practical utility, identifies recurring methodological limitations, and highlights underexplored aspects of performance data. By charting prevailing trends and research gaps, we seek to steer future investigations towards rigorous, generalisable, and practitioner-oriented AI frameworks for objective skill assessment in badminton. To the best of our knowledge, this is the first such review on AI-based badminton player performance assessment in the relevant literature.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cp\u003eThis study follows a systematic literature review methodology guided by the PRISMA framework [20]. The review protocol adapts the procedures outlined by Helbach et al. [21], beginning with the definition of the study\u0026rsquo;s scope and objectives. Specifically, the review addresses three research questions (RQs):\u003c/p\u003e\n\u003cp\u003eRQ1. Which AI approaches are currently applied to badminton player assessment, and what evidence supports their effectiveness?\u003c/p\u003e\n\u003cp\u003eRQ2. What theoretical, methodological, and technical constraints are most frequently reported across these studies?\u003c/p\u003e\n\u003cp\u003eRQ3. How are AI assessment results communicated to ensure they are clear and trusted by coaches and athletes?\u003c/p\u003e\n\u003cp\u003eThese questions structure the subsequent stages of database searching, study screening, data extraction, and synthesis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1. Search Strategy\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe keywords , presented in Table I, were selected to reflect three core themes: the sport context (e.g., \u0026ldquo;badminton\u0026rdquo;, \u0026ldquo;racket sport\u0026rdquo;), type of analysis (e.g., \u0026ldquo;pose estimation\u0026rdquo;, \u0026ldquo;performance assessment\u0026rdquo;), and computational approaches (e.g., \u0026ldquo;machine learning\u0026rdquo;, \u0026ldquo;computer vision\u0026rdquo;). This structure helped capture a comprehensive yet targeted set of publications relevant to the review\u0026rsquo;s objectives.\u003c/p\u003e\n\u003cp\u003eTable I Final search queries used across selected academic databases.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"586\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 33.5043%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDatabase\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66.4957%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFinal query\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 33.5043%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eWeb of Science Core Collection\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66.4957%;\"\u003e\n \u003cp\u003e\u003cem\u003eTS=((badminton OR \u0026quot;racket sport*\u0026quot; OR \u0026quot;racquet sport*\u0026quot;) AND (video analysis OR motion tracking OR pose estimation OR action recognition OR movement analysis OR \u0026quot;performance assessment\u0026quot;) AND (artificial intelligence OR machine learning OR deep learning OR neural network* OR computer vision))\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 33.5043%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eScopus\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66.4957%;\"\u003e\n \u003cp\u003e\u003cem\u003e(badminton OR \u0026quot;racket sport*\u0026quot; OR \u0026quot;racquet sport*\u0026quot;) AND (\u0026quot;video analysis\u0026quot; OR \u0026quot;motion tracking\u0026quot; OR \u0026quot;pose estimation\u0026quot; OR \u0026quot;action recognition\u0026quot; OR \u0026quot;movement analysis\u0026quot; OR \u0026quot;performance assessment\u0026quot;) AND (\u0026quot;artificial intelligence\u0026quot; OR \u0026quot;machine learning\u0026quot; OR \u0026quot;deep learning\u0026quot; OR \u0026quot;neural network*\u0026quot; OR \u0026quot;computer vision\u0026quot;)\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 33.5043%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIEEE Xplore\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66.4957%;\"\u003e\n \u003cp\u003e\u003cem\u003e((\u0026quot;badminton\u0026quot; OR \u0026quot;racket sports\u0026quot; OR \u0026quot;racquet sports\u0026quot;) AND ((\u0026quot;artificial intelligence\u0026quot; OR \u0026quot;machine learning\u0026quot; OR \u0026quot;deep learning\u0026quot; OR \u0026quot;computer vision\u0026quot; OR \u0026quot;neural networks\u0026quot;)) AND ((\u0026quot;player assessment\u0026quot; OR \u0026quot;skill classification\u0026quot; OR \u0026quot;performance metrics\u0026quot; OR \u0026quot;technical analysis\u0026quot; OR \u0026quot;sports analytics\u0026quot;)))\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e2.2. Eligibility Criteria and Study Selection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe initial screening was done by the first and fourth authors. Once the screening was completed, all articles that were not related to the topic were excluded from further consideration. For articles lacking a clear title and abstract, full-text screenings evaluated the content to determine whether they should be included. Decisions about which articles to include were made after careful full-text evaluations, with all authors in full agreement. Figure 1 illustrates the data extraction process, providing a clear overview of how the 356 records retrieved from the academic databases were processed to result in the 51 studies that are included in the data synthesis and analysis of this review.\u003c/p\u003e\n\u003cp\u003eEligibility criteria were established to ensure the relevance of the review to AI-based badminton player assessment. Included studies were peer-reviewed journal articles or conference proceedings published between January 2018 and December 2025 (note: a few early-access articles dated 2026 were also inlcuded) in English. Exclusions applied to abstract-only entries, non-English publications, and studies before 2018. Articles had to focus on AI applications for player assessment, skill evaluation, or performance classification, with methodological clarity and empirical implementation. Studies on tactical analysis, match prediction, or non-AI approaches were excluded unless they addressed player-level assessment. Relevant literature was identified through systematic searches in IEEE Xplore, Scopus, and Web of Science.\u003c/p\u003e"},{"header":"3. Results","content":"\u003cp\u003e3.1. Overview of AI Methods Applied to Badminton\u0026nbsp;Player Assessment\u0026nbsp;(RQ 1)\u003c/p\u003e\n\u003cp\u003e3.1.1. Study Characteristics\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 2 illustrates the yearly publication trend on AI applications for badminton player assessment. Research output was low from 2018 to 2020, with one publication per year. A clear increase can be seen in 2022, rising to ten publications and remaining similar in 2023 and 2024. The number of studies then increased sharply to nineteen in 2025, indicating accelerated research activities in the field. This trajectory indicates growing scholarly interest coinciding with the maturation of deep-learning toolkits that lower technical barriers to entry.\u003c/p\u003e\n\u003cp\u003e3.1.2. Methodological Approaches\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable II. Methodological approaches employed in AI-based badminton player assessment studies.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNo.\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eStudy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eStudy Focus\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePlayer Sample\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eData Source(s)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAI Methodology Applied\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePerformance Metrics Reported\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAssessment Framework Orientation\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation Approach\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[22]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStroke-level dataset for tactical performance prediction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTransformers, Bidirectional-Gated Recurrent Unit (GRU), Graph Models\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eArea Under the Curve (AUC), Accuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTask Benchmarks\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[5]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAutomatic point and stroke annotation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eConvolutional neural network (CNN) and Support Vector Machine (SVM)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy, Mean Average [email protected], Edit Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSplit-based Evaluation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[6]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eProfiling performance via score progression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMatch Data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSVM, RF, K-means\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eR\u0026sup2;, Accuracy, Std. Dev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Test Spl\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[7]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrediction and tactical decision analysis in women\u0026rsquo;s singles\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDT, RF, XGBoost, SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy (up to 87.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCross-validation (90/10 split)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[8]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eClosed-loop AI for imitation, simulation, and strategy optimisation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (ShuttleSet, annotated datasets)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRallyNet, Simulation Environment\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLoss metrices, Win Rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTactical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBenchmarking vs baselines (ShuttleNet and DyMF) in simulated matches\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[9]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003ctable border=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eBehaviour analysis from broadcast videos with visual analytics\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003ctable border=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eYou Only Look Once (YOLO)v3 Model, OpenPose, ResNet-18, LSTM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy (97%), Precision, Recall, F1 (hit-frame detection)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTactical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSplit-based Evaluation (train/val/test; semi-supervised learning)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[10]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy assessment of forehand smash biomechanics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCollegiate athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (front-view rallies, Mediapipe extraction)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSkeletal Pose Estimation (MediaPipe), Dynamic Model, RF, Ridge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy (up to 97.4%), Precision, Recall, F1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Test Split (70/30), Cross-validation, Expert validation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[11]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDeep learning-based recognition of badminton actions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecreational athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHybrid Approach\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLong Short-Term Memory (LSTM), Continuous Learning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy (63%, 84%, 92% across datasets)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBenchmark on HKSR and NTU skeleton datasets\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[12]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDetection and classification of six-corner footwork\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecreational athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eYOLOv8, YOLOv9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMean average precision (0.633 vs 0.605), Precision, Recall, F1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Val/Test Split (80/10/10); Model comparison\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[13]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eShot and strategy recognition using wearable sensors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite and semi-professional athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWearable Sensors\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCNN, LSTM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eShot Accuracy (90.9%), Strategy Accuracy (80%), Precision, Recall, F1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Val/Test Split (64/20/16), Class weighting\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[17]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAction recognition using spatio-temporal skeleton features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eYouth athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (3D skeletal data via MediaPipe)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWeighted Ensemble Machine Learning Models\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy (95.38%), Precision, Recall, F1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5-fold Cross-validation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e12\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[18]\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDetection and classification of six badminton strokes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (Kaggle badminton_stroke_video, 3000+ annotated images)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eYOLO-HGNet\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003emAP (96.1%), Accuracy (95.4%), Precision, Recall, F1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBenchmark vs YOLOv5, OpenPose, CatBoost, XGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[23]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFootwork recognition and 3D trajectory extraction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSub-elite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (binocular cameras)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFaster R-CNN Binocular Positioning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eShoe localisation accuracy (97.2%) and Positioning error (0.129 m)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Val/Test Split (90/10), Trajectory validation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[24]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePlayer evaluation using DRL with technical and tactical contexts\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (BWF match videos, 2018\u0026ndash;2020)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDRL (LSTM-based Q-function), AlphaPose, TrackNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAction Value (Q-values), Correlation with Score/Rank\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBaseline comparisons, Spearman Correlation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[25]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eShot prediction and classification from broadcast videos\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecreational athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (26 broadcast videos, Shuttlecock Trajectory Dataset)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eYOLOv5, DeepSORT, CNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMOTA (73%), HOTA (72%), IDF1 (73%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBenchmark with trajectory metrics\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[26]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eClassification of badminton shots (lob, smash, net)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (BWF Men\u0026rsquo;s Singles matches)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKeras-Mediapipe, YOLO-NAS (Neural Architecture Search)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy (73.5%), Precision, Recall, F1, AUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Val/Test Split (70/10/20), Confusion Matrix\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[27]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePredicting shot accuracy using Quiet Eye and biomechanics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite, Intermediate, and Novice athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHybrid Approach (Eye-tracking + Motion Capture)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNeural Network, SHapley Additive exPlanations (SHAP)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy (85.7%), Precision (88.3%), Recall (83.1%), F1 (0.856), AUC (0.823)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Val/Test Split (70/15/15), Ablation Study\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[28]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCoaching framework for stroke, stance, and scoring\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eProfessional, Intermediate, and Novice athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWearable Sensors (IMU: wrist, palm, left leg, right leg)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCNN (stroke classification, score regression), kNN (stance error estimation)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStroke Accuracy (89.1%), Performance R\u0026sup2; (0.888), Success Rate (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCross-person validation, Train/Val/Test Split (60/20/20), Baseline comparisons\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cspan lang=\"DE\"\u003e[29]\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eShot influence prediction using long/short-term dependencies\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVideo Footage (43,838 shots, 4,350 rallies, 75 matches, 2018\u0026ndash;2021)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCNN (short-term), Transformer (long-term)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAUC (0.871), Brier Score (0.143), Efficiency gains\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5-fold Cross-validation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[30]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSmash detection using optimised deep learning model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVideo Footage (YouTube broadcast matches)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eResidual-Shuffle Net + Upgraded Pufferfish Optimiser (UPO)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAccuracy (96.4%), Precision (0.95), Recall (0.87), F1 (0.87)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTrain/Test Split (85/15), Baseline comparisons (AlexNet, GoogleNet, ResNet-18, I3D)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[31]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePose estimation with improved YOLOv8-Pose and attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecreational athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (xBHPE dataset, Kinect v2, 4000 samples)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eYOLOv8-Pose + Local Attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMean squared error (12.72), Percentage of keypoints @0.2 (0.7793), Frame per second (66.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBenchmark vs HigherHRNet, BlazePose, LitePose\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[32]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBiomechanics, injury prevention, and tactical analysis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid Approach (Video, Wearable Sensors, Historical Match Data)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCNN (pose estimation)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCV Models (93% accuracy), Injury Prediction (90%), RL Defence Gain (20%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTechnical and Tactical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCross-validation, Baseline comparisons, Empirical training tests\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[33]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAutomated detection of badminton players from broadcast videos\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVideo Footage (YouTube broadcast matches: 2011 All England, 2012 Olympics, 2017 Asia Championship)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eFaster R-CNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAverage Precision (PR curve-based)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCase-based training/testing on single and combined videos\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[34]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAutonomous motion tracking of squash players\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVideo Footage (broadcast squash matches, PSA Canary Wharf Classic 2013)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eMulti-person Pose Estimation (CNN),\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eR\u0026sup2; (0.99), Error in Distance (3.73%), Speed Error (1.26%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBenchmark vs manual tracking, Filtering vs Unfiltered data\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[35]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3D shuttle trajectory reconstruction from monocular video\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (TrackNetV2 dataset + 40 YouTube matches)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eGraph-based court detection, U-Net, GRU\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHit Detection (94.6% F1), 3D Reconstruction Error (~8 cm synthetic)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBenchmarks on TrackNetV2 Synthetic Trajectory Evaluation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[36]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSurvey of video action recognition in sports (datasets, methods, applications)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMultiple sports (team \u0026amp; individual)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (SoccerNet, Badminton Olympic, Diving48, FineGym, TTStroke-21, etc.)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2D/3D CNN, Two-stream, Transformer, Skeleton-based\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cem\u003e-\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eComparative analysis across datasets and models\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[37]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCourt line extraction from broadcast badminton videos\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVideo Footage (BadmintonWorld.tv tournament videos)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHorizontal Line Projection + K-means, Morphological Corner Detection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePixel Error (0\u0026ndash;3 px), Frame Rate (33.3 FPS)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eComparative benchmark vs Hough Transform methods\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[38]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMotion trajectory tracking with improved KCF and depth fusion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (binocular cameras, badminton arm strokes)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eViBe (background subtraction), KCF + Depth Fusion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTracking Success Rate (98%), Trajectory Error (0.365), Speed Error (0.116)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBenchmark vs Kalman Filter \u0026amp; Original KCF\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[39]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStroke recognition in badminton using TSN with attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (badminton stroke videos)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTemporal Segment Network (TSN) + Lightweight Attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall (91.2%), Accuracy (91.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Test Split\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[40]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eShot refinement by combining shuttle tracking and hit detection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (BWF matches, 32 + 354 matches, 8,975 + 98,675 frames)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrackNet (shuttle), YOLOv7 (swing), DensePose (pose), Shot Refinement Algorithm (SRA),\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eShot Detection: Precision 0.897, Recall 0.913, F1 0.905; Shot Classification: Accuracy 72.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBenchmark vs TrackNet; Train/Val/Test split; t-IoU evaluation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[41]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eShuttlecock detection with lightweight small-object detector\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (custom shuttlecock dataset, 6113 train / 1586 test images)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eYOLO-inspired CNN\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMean Average Precision (98% shuttlecock), FPS (30 on Jetson Nano)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBenchmark vs YOLOv3/v4/v5, v3-tiny, v4-tiny on public and shuttle datasets\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[42]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRally outcome prediction using stroke sequences\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideo Footage (London 2012 Olympics, 10 annotated players, 498 rallies)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eResNet-18, Bidirectional LSTM, Faster CNN\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy (0.70), Brier Score (0.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Test Split (398/100 rallies)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[43]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eStroke detection using wearable wrist device\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eRecreational athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eWearable Sensors (Arduino Nano 33 BLE, accelerometer, gyroscope)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCNN (embedded), Motion segmentation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAccuracy (100% in small dataset), Confusion Matrix\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTechnical; Real-time\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePrototype testing on limited stroke dataset\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[44]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOptimisation of doubles positioning and movement\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMatch Data (2019 BWF Super Series finals, 10 matches)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNumerical Modeling (7 nonlinear equations, MATLAB fsolve)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOptimal attack \u0026amp; defense zones, Running circle model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSimulation and validation with BWF match statistics\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[45]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKinect based posture recognition and action evaluation system\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBadminton athletes and collegiate participants\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMulti Kinect V2 skeletal and depth data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSensor calibration, Kalman filtering, CNN with attention, Transformer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy up to 98.75%; MAE 0.1364; F1 score 0.6911; 9.27 ms processing time\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBaseline comparison and ablation study\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[46]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVision-based badminton stroke recognition and tactical analysis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite badminton athletes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBroadcast and match video footage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eImproved YOLO-based detection, pose estimation, temporal sequence modelling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStroke classification accuracy above 90%; mAP and F1-score reported\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Test split, baseline comparison with standard YOLO variants\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[47]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVision-based badminton posture recognition and training assistance system\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBadminton players (experimental evaluation cohort)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eImproved OpenPose (MobileNet backbone) + Particle Filter tracking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eImproved OpenPose with MobileNet backbone; Particle Filter tracking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFPS, Recall, Precision, Accuracy, F1-score, response time, CPU usage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eExperimental evaluation and model comparison\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[48]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHybrid vision\u0026ndash;language framework (ChatMatch) for badminton video understanding\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eProfessional players and coach (evaluation); elite match videos\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBroadcast match videos (20 singles matches)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUNet (court segmentation), YOLOv5, ResNet-50 (action recognition), GPT-3.5-based multi-agent system\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLocation accuracy 0.991; Action accuracy 0.902; Gesture accuracy 0.950\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Test split; confusion matrices; user-based evaluation.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[49]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eIntelligent badminton handle with multinode MEMS sensors for explainable motion recognition\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBadminton players (experimental evaluation cohort)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMultinode MEMS sensors embedded in racket handle (accelerometer and gyroscope data)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCNN-based motion recognition with explainability analysis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecognition accuracy above 90%; precision, recall, F1-score reported\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eExperimental validation with controlled stroke trials; comparative evaluation with baseline models\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[50]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eInertial sensor\u0026ndash;based badminton swing recognition\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e10 players; 1,200 swing samples\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRacket-mounted accelerometer and gyroscope data; VideoBadminton dataset\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePCA, SVM (grip classification), AdaBoost (six-class swing recognition)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy 95.93%; Precision, Recall, F1-score; AUC up to 0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5- and 10-fold cross-validation; baseline comparison\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[51]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eGrip-force analysis for badminton performance evaluation using flexible pressure sensors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e30 male participants (15 beginners, 15 national-level athletes)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFlexible piezoresistive pressure sensors embedded in racket handle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSignal acquisition and force-feature extraction (Imax, Iave, T); statistical correlation analysis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eGrip peak force, mean force, duration; performance scores; correlation coefficients (r \u0026asymp; \u0026minus;0.6 to \u0026minus;0.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eGroup comparison (beginner vs athlete); correlation analysis; significance testing (p \u0026lt; 0.05)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[52]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDeep learning and analytics for biomechanics, injury prediction, and tactical modelling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCompetitive badminton players\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMatch video, wearable sensors, historical data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCNN pose estimation; Random Forest (injury).\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e93% biomechanical accuracy; 90% injury prediction; 15% speed gain; 25% injury reduction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eControlled evaluation, confusion matrix, tactical performance comparison\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[53]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eMotion tracking and decomposition model for badminton training\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBadminton training participants\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3D limb joint coordinate data via motion sensing system\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3D decomposition model; Gauss-Newton optimisation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy up to 99.31%; \u0026gt;34% improvement vs LSTM; 32.5% faster response\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBaseline comparison with LSTM; model accuracy and efficiency tests\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[54]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVision-based badminton stroke detection and performance evaluation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCompetitive badminton players\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMatch video footage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDeep CNN-based detection and temporal sequence modelling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy, Precision, Recall, F1-score reported\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eExperimental evaluation and baseline comparison\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[55]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eIntegrated match analysis: detection, tracking, shot classification\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite match videos\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBroadcast footage; VideoBadminton dataset\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eModified YOLOv8; court-aware tracking; SlowFast + TimeSformer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\[email protected] 94.12%; Top-1 77.8%; Top-5 94.24%; 28 FPS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTrain/Test split; baseline comparison; confusion matrix\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[56]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eReal-time action recognition and pose estimation with shot refinement\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eProfessional match videos (62 matches)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBroadcast footage with annotated hits\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eImproved YOLOv8 + 13-keypoint pose model + Kalman filtering\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy 90.2%; F1-score 87.5%; Temporal error 2.14 frames\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical and Tactical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBaseline comparison; cross-scenario evaluation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[57]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eStroke recognition using Quantum CNN (QCNN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCompetitive players (540 stroke samples)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBroadcast videos; OpenPose joint data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eQCNN vs SVM and 3D-CNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy and F1 up to 0.965; noise robustness ACC 0.882\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e70/15/15 split; baseline comparison; noise evaluation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[58]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDeep learning\u0026ndash;based badminton action recognition with multimodal fusion and computational optimisation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e19 university-level athletes (VideoBadminton dataset; 7,822 clips)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eVideoBadminton dataset (annotated match videos)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3D-CNN, Bi-LSTM, Video Swin Transformer; multimodal fusion; transfer learning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy up to 0.967; F1 up to 0.953; inference time 10.8\u0026ndash;13.1 ms; reduced FLOPs and memory usage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e70/15/15 split; comparison with temporal graph network and other baseline models; efficiency and performance benchmarking\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[59]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBadminton action recognition with multimodal fusion and computational optimisation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e19 university athletes (VideoBadminton dataset; 7,822 clips; 18 action classes)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAnnotated match videos (VideoBadminton)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3D-CNN, Bi-LSTM, Video Swin Transformer; multimodal fusion; transfer learning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy up to 0.967; F1 up to 0.953; MSE 0.024\u0026ndash;0.037; inference 10.8\u0026ndash;13.1 ms\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e70/15/15 split; comparison with temporal graph convolutional network and other multimodal networks; efficiency benchmarking\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[60]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKey frame detection in badminton swings for educational feedback\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUniversity students (102 videos; 143 in intervention)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTablet video; MediaPipe 3D skeleton data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eST-GCN; MLP; temporal correction algorithm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall; F1-score; Average Positional Error\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTechnical; Real-time and Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCross-validation; intervention study; statistical evaluation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[3]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTactical badminton analysis using computer vision and Retrieval-Augmented Generation (RAG)-anhanced large language model (LLM)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eElite singles matches (\u0026gt;750 rallies)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1080p broadcast videos (BWF)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eYOLOv8; MediaPipe pose; multi-cue shot detection; RAG-LLM (LangChain + FAISS)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003emAP 0.94 (player); 0.83 (shuttle); F1 0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTactical; Post-hoc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eQuantitative evaluation; statistical testing; expert validation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable II shows the range of research designs in AI-based badminton player assessment. Most studies employ controlled experiments to analyse technical execution, biomechanics, and tactics, though often with reduced ecological validity. Observational approaches using broadcast or archived footage capture authentic match play but are limited by video quality. A smaller number of works focus on algorithmic innovation or feasibility studies addressing deployment challenges. The table also shows a concentration on elite and collegiate athletes, with fewer studies on recreational or youth groups, such as Wang et al. [9] \u0026nbsp;for elite tactical analysis, Sinadia and Murwantara [6] for collegiate profiling, Liu and Liang [11] for recreational cohorts, and Amudhan et al. [41] for youth shuttlecock detection. This emphasis provides insights for high-performance contexts but limits generalisability to the wider badminton community, which is dominated by recreational and developing players [28,41,43].\u003c/p\u003e\n\u003cp\u003eMost AI-based badminton player assessment studies rely on video footage as the primary data source [3,5,7\u0026ndash;9], reflecting its accessibility through broadcasts and archives and its minimal burden on players. A smaller set employs wearable sensors to capture biomechanical \u0026nbsp;or physiological data [51], with only a few combining video and sensors to leverage spatial\u0026ndash;temporal resolution alongside on-body kinematics. For instance, Wang et al. [9] introduced the ShuttleSet dataset with annotated video for tactical analysis, Van Herbruggen et al. [13] integrated inertial and ultra-wideband sensors with video to study match strategies, and Zheng and Chen [32] fused video, sensors, and historical records to support both tactical assessment and injury prevention. While video remains the most practical modality, hybrid approaches offer promising opportunities to enhance accuracy and contextual richness in future badminton analytics.\u003c/p\u003e\n\u003cp\u003e3.1.3. AI Methodologies for Badminton Player Assessment\u003c/p\u003e\n\u003cp\u003eAI research in badminton can be categorised into four principal application areas: stroke analysis, movement-pattern analysis, spatio-temporal or tactical analysis, and multi-modal technical evaluation. The literature is predominantly concentrated on stroke-level modelling, where computer vision, deep learning, and pose estimation techniques are applied to detect and classify strokes from video data. Movement-pattern studies focus on footwork recognition, trajectory tracking, and positional dynamics to characterise court coverage and recovery behaviour. Spatio-temporal or tactical approaches examine rally sequences, shot influence, and strategic decision-making processes, extending analysis beyond isolated actions to match-level contexts. Multi-modal technical evaluation integrates heterogeneous data sources, including video, wearable sensors, and biomechanical signals, to provide more comprehensive performance assessment. Overall, current research emphasises visual stroke detection, while integrated and strategy-oriented evaluation frameworks remain comparatively less developed.\u003c/p\u003e\n\u003cp\u003eStroke analysis tasks in badminton frequently adopt deep learning classifiers such as 2D- CNNs, 3D-CNNs, and transformer-based sequence models to recognise stroke categories including smashes, drops, clears, and drives [5,18,30,39]. CNN-based approaches excel at capturing spatial features from individual frames, while 3D-CNNs extend this to spatio-temporal volumes, improving recognition of subtle motion cues. Transformer architectures and temporal CNNs further enhance performance by modelling long-range dependencies across rally sequences, making them effective in distinguishing strokes with similar visual appearance but different execution timing. Reported accuracies often exceed 85\u0026ndash;90%, highlighting the maturity of these models for technical skill classification \u0026nbsp;[10,13,17,18,23,29]. Movement-pattern analysis, on the other hand, builds on pose estimation outputs from frameworks such as OpenPose and MediaPipe, which generate joint keypoints used as input features for downstream models [10\u0026ndash;12,31]. Recurrent neural networks, long short-term memory (LSTM), and graph convolutional networks are commonly employed to capture temporal dependencies and relational dynamics between body joints, enabling assessment of agility, footwork recovery, and tactical positioning [22]. Some studies incorporate You Only Look Once (YOLO)-based detectors for robust player localisation before pose extraction, ensuring consistent tracking under broadcast video conditions [9,12,18,28,31].\u003c/p\u003e\n\u003cp\u003eMovement-pattern and rally-sequencing analyses in badminton performance span spatial, temporal, and cognitive domains, including court coverage [22,48], footwork sequencing [12,23], recovery positioning [10], and contextual rally dynamics [7\u0026ndash;9]. Spatial mapping and trajectory overlays produce heat-map visualisations aligned with expert evaluations [12,23], while pose-estimation models reliably classify six-corner footwork patterns [9]. Recovery-position models based on skeletal features achieve high accuracy but vary in predictive stability, reflecting the situational complexity of tactical decision-making [10]. Rally-sequencing studies highlight strategic shot-to-shot transitions that distinguish playing styles and outcomes [7\u0026ndash;9], with machine learning approaches including decision tree (DT), random forest (RF), and graph models used to evaluate shot selection in rally contexts [7,22]. Integrating perceptual\u0026ndash;cognitive markers, Tan and Teoh [27] combined Quiet Eye metrics with biomechanical features in a neural network, achieving over 85% accuracy in predicting shot outcomes and identifying gaze duration and onset timing as critical predictors.\u003c/p\u003e\n\u003cp\u003e3.1.4. Assessment Validation Approaches\u003c/p\u003e\n\u003cp\u003eAI-based badminton player assessment employs four main categories of performance indicators: biomechanical, shot quality, tactical, and movement-related metrics. Biomechanical measures, such as joint angles and segment velocities, are validated against expert annotations and biomechanical reference models, and these features are subsequently encoded as kinematic vectors for classifier training [10,12,40]. Shot-quality indicators, including accuracy, consistency, and power, are benchmarked using supervised classification pipelines with cross-validation, and they function as target variables in stroke recognition models [26]. Tactical metrics, such as shot selection and court positioning, are assessed through predictive accuracy and correlations with rally outcomes [55], and they are typically modelled as sequential dependencies within recurrent or transformer-based architectures [25]. Movement-related measures, notably footwork efficiency and court coverage, are validated using pose-tracking precision and trajectory error, and they are structured into graph or temporal representations for learning locomotor strategies [12].\u003c/p\u003e\n\u003cp\u003eWithin AI-based badminton assessment, existing frameworks vary according to their analytical focus and temporal orientation. Most are designed as technical post-analysis systems, emphasising stroke mechanics and biomechanical precision once matches have concluded [7,17]. A smaller body of work addresses real-time technical assessment, where immediate feedback is achieved under favourable computational conditions [40]. Tactical strategy frameworks are reported less frequently and are generally retrospective, analysing rally sequences and decision-making behaviour through post-match data [8]. Real-time tactical guidance systems are rare, though studies demonstrate their capacity to provide adaptive decision support during active play [9,12,23,38]. Hybrid frameworks that combine technical and tactical elements across flexible timelines are limited but illustrate attempts to capture mechanical precision alongside contextual dynamics [24,28].\u003c/p\u003e\n\u003cp\u003e3.2. Common Challenges in AI-Based Badminton Assessment Research (RQ 2)\u003c/p\u003e\n\u003cp\u003eA recurrent theme across the reviewed literature is the persistence of technical challenges that constrain the reliability and scalability of AI-based badminton assessment systems. Data acquisition difficulties are the most frequently reported, particularly in relation to shuttlecock tracking and court reference extraction. Hsu et al. [40] highlighted the limitations of monocular video, where occlusion and high shuttle speed hinder consistent localisation, thereby impairing downstream analytical processes. Similarly, Wei and Weng [37] demonstrated that court line extraction from broadcast footage remains sensitive to lighting variation and camera zoom, reducing the stability of detection pipelines. These examples underscore how the quality and consistency of raw input data continue to represent a major obstacle for performance assessment frameworks.\u003c/p\u003e\n\u003cp\u003eIntegration complexity emerges as another common limitation, particularly in systems that combine multiple data modalities. Ghosh et al. [28] emphasised the engineering burden of synchronising heterogeneous streams within the DeCoach framework, while Van Herbruggen et al. [13] noted similar challenges when aligning inertial sensor data with high-frame-rate video to maintain biomechanical fidelity. Closely related are concerns of generalisability, as many models perform well within narrowly defined datasets but degrade when applied to different playing styles or populations. Baclig et al. [34] demonstrated this difficulty when multi-player tracking models developed in squash contexts lost accuracy in badminton rallies, and Zhi et al. [39] reported analogous domain-shift effects in stroke recognition tasks.\u003c/p\u003e\n\u003cp\u003eFurther limitations involve computational efficiency, real-time processing, and tracking precision. Yang et al. [18] explored model pruning and mixed-precision computation to preserve analytic accuracy while remaining within hardware constraints, whereas Salim et al. [26] designed adaptive inference pipelines that modulate processing load according to rally tempo. Tracking accuracy has also been identified as a recurring difficulty, as even minor errors in court-line detection can propagate across extended rallies, producing cumulative distortions in positional analysis [38,39]. These technical challenges illustrate the ongoing tension between methodological sophistication, practical feasibility, and ecological validity in AI-based badminton performance assessment.\u003c/p\u003e\n\u003cp\u003e3.3 Interpretability and Usability of AI Outputs in Badminton Player Assessment (RQ3)\u003c/p\u003e\n\u003cp\u003e3.3.1\u0026ensp;Model Interpretability and Explainable Vision\u0026ndash;Language Frameworks\u003c/p\u003e\n\u003cp\u003eOnly 4 of the 51 reviewed studies incorporated explicit interpretability mechanisms. Wang [8] provided frame-level predictions with Grad-CAM heat maps that highlight the court regions influencing each decision, whereas Chen et al. [9] attached SHAP value attributions to rally-phase classifications. These visual or feature-based explanations transform otherwise opaque predictions into reasoned assessments that align with coaching terminology. Despite their demonstrated value, the limited uptake indicates a substantial gap between algorithmic precision and user confidence.\u003c/p\u003e\n\u003cp\u003eMore recently, interpretability has also been extended beyond feature attribution towards structured vision\u0026ndash;language reasoning. The ChatMatch framework [48] integrates meta-feature extraction, rule-based knowledge decoding, and LLM agents to generate structured and unstructured explanations of match events. By transforming spatial, action, and gesture recognition outputs into descriptive rally narratives and statistical summaries, the system provides traceable reasoning pathways between visual evidence and analytical conclusions [3]. This hybrid architecture illustrates a shift from post hoc visualisation towards interactive, explainable inference mechanisms that support professional query-based analysis.\u003c/p\u003e\n\u003cp\u003e3.3.2\u0026ensp;Implications for Coaching Practice\u003c/p\u003e\n\u003cp\u003eEvidence from practice-focused evaluations shows that interpretable outputs improve both usability and acceptance. DeCoach [28] is used to quantify vertical growth stages objectively. Wang et al. [29] integrated bounding-box projections from their YOLO-HGNet detector with kinematic graphs, allowing for inspection of fast exchanges that are difficult to analyse visually. At larger scales, the ShuttleSet framework introduced by Wang et al. [22] processed more than thirty-six thousand strokes and provides dashboard summaries that shorten the routine assessment of approximately five hundred players. In a complementary line of work, Zhang and Zhong [45] developed a Kinect-based auxiliary training system that integrates multi-sensor skeletal fusion, gesture recognition, and hierarchical action evaluation to support structured badminton teaching. Their system demonstrated high motion-recognition accuracy and enabled real-time corrective feedback, illustrating how sensor-driven pose analysis can be embedded into pedagogical workflows.\u003c/p\u003e\n\u003cp\u003e3.3.3\u0026ensp;Applications for Competition Organisation\u003c/p\u003e\n\u003cp\u003eExplainable assessment outputs also facilitate fairer competition structures. Sinadia and Murwantara [6] demonstrated that machine-learning profiles linked to positional heat maps can refine club assessment systems, while Asriani et al. [17] proved that spatio-temporal rally analysis is more sensitive than fixed-period reviews when reclassifying players. Xie et al. [44] further reported that pose-aware analytics enable organisers to design tournaments tailored to diverse tactical styles. Collectively, these studies suggest that transparent analytics support equitable seeding, responsive regrading, and data-informed scheduling.\u003c/p\u003e\n\u003cp\u003e3.3.4\u0026ensp;Player Development Implications\u003c/p\u003e\n\u003cp\u003eFor athletes, interpretable feedback promotes self-directed improvement. A study by Liu and Liang [11] presents an action-recognition tool that players use between formal sessions to monitor technique independently. Yuan et al. [7] proposed a multi-feature framework that blends technical and tactical indices to personalise training priorities. Tan and Teoh [27] combined neural network shot predictions with accurate metrics, providing objective evidence of skill progression that can enhance motivation. These findings indicate that accessible explanations can extend the impact of AI systems beyond coaching sessions, supporting continuous learning and goal-oriented practice.\u003c/p\u003e\n\u003cp\u003e3.3.5\u0026ensp;Methodological Effectiveness\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMany of the studies reviewed highlight recurring themes in performance testing that are increasingly relevant to practical coaching. One area with strong potential is the use of pose estimation to analyse player movements. For example, the enhanced YOLO model developed by Yang et al. [18] improved the accuracy of posture detection in badminton players by more than eight percent compared to earlier systems. This improvement was achieved by adjusting how the model processes visual information, resulting in greater efficiency and reliability. However, the availability of badminton-specific datasets remains limited, particularly for multi-view capture and synchronised multi-modal data [23], and addressing these constraints is essential for achieving more robust and generalisable pose-estimation models.\u003c/p\u003e\n\u003cp\u003eIn addition to technical evaluation, some studies have explored the use of spatial and temporal features to support player assessment, including measurements of body position and timing. These models offer promising frameworks for translating movement data into meaningful performance insights. However, evaluating tactical decision-making remains more challenging. As shown in the ShuttleSet study by Wang [22], analysing strategic behaviour requires substantial human annotation and expert interpretation. This makes tactical assessment harder to scale and apply consistently. These limitations point to the need for stronger collaboration between AI developers and coaching practitioners to create systems that are not only accurate but also practical and intuitive for use in real sport settings.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThis review set out to clarify three questions: which AI techniques work best for badminton player assessment, what methodological hurdles still stand in the way, and how clearly current systems communicate their findings. Thirty-four primary studies, summarised in Table II, show that computer vision pipelines now dominate the field. Yet, they sit alongside movement-pattern analytics, spatio-temporal sequence models, and a small but important group of multi-modal systems. Together, these works sketch a discipline that has matured rapidly but remains uneven in scope and depth.\u003c/p\u003e \u003cp\u003eModern vision models that track body landmarks and shuttle flight now achieve remarkable accuracy, with Yang et al. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] reporting stroke-recognition rates above 95%, demonstrating rapid progress in automated analysis. By combining image cues with temporal contexts, these systems can assess not only what players do but also when and why actions matter within a rally, providing coaches with quantitative insights into technique, tactical rhythm, and fatigue. Yet, despite matching or exceeding human reliability in routine evaluations, many studies acknowledge that performance often degrades under varying recording conditions, highlighting the gap between technical promise and practical robustness and raising the critical question of what barriers limit broader adoption.\u003c/p\u003e \u003cp\u003eData collection emerges as the most frequently reported obstacle in AI-based badminton player assessment. Hsu et al. [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e] observed that the shuttlecock often disappears from a single camera\u0026rsquo;s field of view due to its high velocity, disrupting otherwise reliable analytic pipelines. Integration complexity represents another recurrent challenge, with Ghosh et al. [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] documenting the substantial effort needed to synchronise video streams, inertial sensors, and match statistics within the DeCoach framework. Additional constraints arise from limited generalisability, computational overhead, and real-time latency, all of which highlight the contrast between controlled laboratory benchmarks and the variable conditions of competition environments characterised by fluctuating lighting, multiple camera angles, and heterogeneous player styles.\u003c/p\u003e \u003cp\u003eOnly two studies, Wang [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] and Chen et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], integrated explanation tools such as Grad-CAM or SHAP, showing that even simple visual cues boost coach confidence and player engagement. More recent systems, including ChatMatch [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e] and Court-to-Conversation [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], integrated RAG and LLM components to transform detected visual events into structured, query-driven analytical feedback. By grounding language outputs in retrieved match data and recognised actions, these frameworks establish traceable links between raw visual evidence and narrative explanations, thereby supporting coach-player interaction. Future progress depends on open multi-modal datasets, models with domain adaptation and uncertainty estimation, and interpretability as a standard rather than a novelty. Longitudinal studies that track how AI-feedback shapes skill development and performance will then provide the evidence needed for widespread adoption.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis systematic review examined the current development of AI applications in badminton player assessment, highlighting both methodological diversity and several key limitations. Although AI systems offer the potential for objective, consistent, and multi-dimensional performance evaluation, their adoption remains limited due to weak theoretical grounding, insufficient methodological rigour, and limited contextual adaptability. Existing studies employ various approaches, including computer vision for stroke recognition, movement analysis, spatio-temporal modelling, and multi-modal frameworks. Future research should strengthen the theoretical foundation by linking computational outputs with established models of skill acquisition and athlete development, while also developing standardised performance metrics to enable meaningful comparisons across studies. A promising future trajectory lies in developing interactive and explainable AI systems supported by LLMs, enabling performance analysis to be translated into natural, context-aware explanations that coaches and players can easily interpret and utilise.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eCompeting Interests\u003c/h2\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial or non-financial interests that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThe authors received no specific funding for this work.\u003c/p\u003e\n\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\n\u003cp\u003eR.C. and F.U.D. designed the study and supervised a student group in carrying out the initial analysis. C.A. carried out further analysis and wrote the manuscript. All authors reviewed, edited and approved the final manuscript.\u003c/p\u003e\n\u003ch2\u003eAcknowledgement\u003c/h2\u003e\n\u003cp\u003eThe authors would like to acknowledge the initial work carried out by a student group led by Utsab Gyawali, which contributed to the early exploration of this research topic.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eW\u0026ouml;rner EA, Safran MR. Racquet Sports: Tennis, Badminton, Racquetball, Squash. Specific Sports-Related Injuries, Cham: Springer International Publishing; 2021, p. 431\u0026ndash;46. https://doi.org/10.1007/978-3-030-66321-6_30.\u003c/li\u003e\n\u003cli\u003eWoo T-MT, Alam F. Comparative aerodynamics of synthetic badminton shuttlecocks. Sports Engineering 2018;21:21\u0026ndash;9. https://doi.org/10.1007/s12283-017-0241-2.\u003c/li\u003e\n\u003cli\u003eBharadwaj K, Srinivasa G. Court to conversation: Tactical badminton analysis via computer vision and RAG-enhanced LLMs. Knowl Based Syst 2026;333:115027. https://doi.org/10.1016/j.knosys.2025.115027.\u003c/li\u003e\n\u003cli\u003eJekauc D, Burkart D, Fritsch J, Hesenius M, Meyer O, Sarfraz S, et al. Recognizing affective states from the expressive behavior of tennis players using convolutional neural networks. Knowl Based Syst 2024;295:111856. https://doi.org/10.1016/j.knosys.2024.111856.\u003c/li\u003e\n\u003cli\u003eGhosh A, Singh S, Jawahar C V. Towards structured analysis of broadcast badminton videos. Proceedings of the IEEE Winter Conference on Applications of Computer Vision 2018:296\u0026ndash;304. https://doi.org/10.1109/WACV.2018.00039.\u003c/li\u003e\n\u003cli\u003eSinadia HE, Murwantara IM. Sports analytics: A comparison of machine learning performance for profiling badminton athlete. Proceedings of the International Conference on Technology Innovation and Its Applications 2022. https://doi.org/10.1109/ICTIIA54654.2022.9935852.\u003c/li\u003e\n\u003cli\u003eYuan H, Wang Y, Yang K, Bin Y. Prediction model and technical and tactical decision analysis of women\u0026rsquo;s badminton singles based on machine learning. PLoS One 2024;19. https://doi.org/10.1371/journal.pone.0312801.\u003c/li\u003e\n\u003cli\u003eWang KD. Enhancing badminton player performance via a closed‑loop AI approach: Imitation, simulation, optimization, and execution. Proceedings of the International Conference on Information and Knowledge Management 2023:5189\u0026ndash;92. https://doi.org/10.1145/3583780.3616001.\u003c/li\u003e\n\u003cli\u003eChen WT, Wu HY, Shih YA, Wang CC, Wang YS. Exploration of player behaviours from broadcast badminton videos. Computer Graphics Forum 2023;42:267\u0026ndash;79. https://doi.org/10.1111/cgf.14786.\u003c/li\u003e\n\u003cli\u003eKrishnaram N, Ahamed J, Sathyamoorthy N, Sandaruwan KD, Athapaththu AMKB. Skeletal point analysis to determine the accuracy of forehand smash shots played by badminton players. J Natl Sci Found 2024;52:125\u0026ndash;42. https://doi.org/10.4038/jnsfsr.v52i1.12141.\u003c/li\u003e\n\u003cli\u003eLiu J, Liang B. An action recognition technology for badminton players using deep learning. Mobile Information Systems 2022. https://doi.org/10.1155/2022/3413584.\u003c/li\u003e\n\u003cli\u003eJannet K, Sunar MS, Molla MMI, Bin As\u0026rsquo;Ari MA. A deep learning approach to badminton player footwork detection based on YOLO models: A comparative study. Proceedings of the IEEE International Conference on Signal and Image Processing Applications 2024. https://doi.org/10.1109/ICSIPA62061.2024.10686537.\u003c/li\u003e\n\u003cli\u003eVan Herbruggen B, Fontaine J, Simoen J, De Mey L, Peralta D, Shahid A, et al. Strategy analysis of badminton players using deep learning from IMU and UWB wearables. Internet of Things 2024;27. https://doi.org/10.1016/j.iot.2024.101260.\u003c/li\u003e\n\u003cli\u003eMarques A, Gouveira \u0026Eacute;R, Peralta M, Martins J, Venturini J, Henriques-Neto D, et al. Cardiorespiratory fitness and telomere length: a systematic review. J Sports Sci 2020;38:1690\u0026ndash;7. https://doi.org/10.1080/02640414.2020.1754739.\u003c/li\u003e\n\u003cli\u003eZhou D, Keogh JWL, Ma Y, Tong RKY, Khan AR, Jennings NR. Artificial intelligence in sport: A narrative review of applications, challenges and future trends. J Sports Sci 2025:1\u0026ndash;16. https://doi.org/10.1080/02640414.2025.2518694.\u003c/li\u003e\n\u003cli\u003eVickery-Howe DM, Bonanno DR, Dascombe BJ, Drain JR, Clarke AC, Hoolihan B, et al. Physiological, perceptual, and biomechanical differences between treadmill and overground walking in healthy adults: A systematic review and meta-analysis. J Sports Sci 2023;41:2088\u0026ndash;120. https://doi.org/10.1080/02640414.2024.2312481.\u003c/li\u003e\n\u003cli\u003eAsriani F, Azhari A, Wahyono W. Improving badminton action recognition using spatio‑temporal analysis and a weighted ensemble learning model. Computers, Materials and Continua 2024;81:3079\u0026ndash;96. https://doi.org/10.32604/cmc.2024.058193.\u003c/li\u003e\n\u003cli\u003eYang W, Jiang M, Fang X, Shi X, Guo Y, Al‑qaness MAA. A high‑precision and efficient method for badminton action detection in sports using You Only Look Once with Hourglass Network. Eng Appl Artif Intell 2024;137:109177. https://doi.org/10.1016/j.engappai.2024.109177.\u003c/li\u003e\n\u003cli\u003eCao Z, Liao T, Song W, Chen Z, Li C. Detecting the shuttlecock for a badminton robot: A YOLO based approach. Expert Syst Appl 2021;164:113833. https://doi.org/10.1016/j.eswa.2020.113833.\u003c/li\u003e\n\u003cli\u003ePage MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021:n71. https://doi.org/10.1136/bmj.n71.\u003c/li\u003e\n\u003cli\u003eHelbach J, Hoffmann F, Pieper D, Allers K. Reporting according to the preferred reporting items for systematic reviews and meta-analyses for abstracts (PRISMA-A) depends on abstract length. J Clin Epidemiol 2023;154:167\u0026ndash;77. https://doi.org/10.1016/j.jclinepi.2022.12.019.\u003c/li\u003e\n\u003cli\u003eWang WY, Huang YC, Ik TU, Peng WC. ShuttleSet: A human‑annotated stroke‑level singles dataset for badminton tactical analysis. Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining 2023:5126\u0026ndash;36. https://doi.org/10.1145/3580305.3599906.\u003c/li\u003e\n\u003cli\u003eLuo J, Hu Y, Davids K, Zhang D, Gouin C, Li X, et al. Vision‑based movement recognition reveals badminton player footwork using deep learning and binocular positioning. Heliyon 2022;8. https://doi.org/10.1016/j.heliyon.2022.e10089.\u003c/li\u003e\n\u003cli\u003eDing N, Takeda K, Fujii K. Deep reinforcement learning in a racket sport for player evaluation with technical and tactical contexts. IEEE Access 2022;10:54764\u0026ndash;72. https://doi.org/10.1109/ACCESS.2022.3175314.\u003c/li\u003e\n\u003cli\u003eAshfaq F, Jhanjhi NZ, Khan NA. Badminton player\u0026rsquo;s shot prediction using deep learning. Lecture Notes in Bioengineering 2023:233\u0026ndash;43. https://doi.org/10.1007/978-981-99-0297-2_19.\u003c/li\u003e\n\u003cli\u003eSalim MA, Williamdy W, Steven E. AI‑powered badminton shot classification. Proceedings of the International Conference on Technology Innovation and Its Applications 2024. https://doi.org/10.1109/ICTIIA61827.2024.10761547.\u003c/li\u003e\n\u003cli\u003eTan S, Teoh TT. Predicting shot accuracy in badminton using quiet eye metrics and neural networks. Applied Sciences 2024;14:9906. https://doi.org/10.3390/app14219906.\u003c/li\u003e\n\u003cli\u003eGhosh I, Ramasamy Ramamurthy S, Chakma A, Roy N. DeCoach: Deep learning‑based coaching for badminton player assessment. Pervasive Mob Comput 2022;83. https://doi.org/10.1016/j.pmcj.2022.101608.\u003c/li\u003e\n\u003cli\u003eWang WY, Chan TF, Peng WC, Yang HK, Wang CC, Fan YC. How is the stroke? Inferring shot influence in badminton matches via long short‑term dependencies. ACM Trans Intell Syst Technol 2022;14. https://doi.org/10.1145/3551391.\u003c/li\u003e\n\u003cli\u003ePeng W, Zheng W. An AI-based badminton smash detection using residual-shuffle network optimized based on upgraded pufferfish optimizer. Ain Shams Engineering Journal 2025;16:103414. https://doi.org/10.1016/j.asej.2025.103414.\u003c/li\u003e\n\u003cli\u003eWu Y, Chen Z, Zhang H, Yang Y, Yi W. Enhanced Pose Estimation for Badminton Players via Improved YOLOv8-Pose with Efficient Local Attention. Sensors 2025;25:4446. https://doi.org/10.3390/s25144446.\u003c/li\u003e\n\u003cli\u003eZheng Y, Chen E. Harnessing deep learning and advanced analytics to revolutionise badminton performance. International Journal of Information and Communication Technology 2025;26. https://doi.org/10.1504/IJICT.2025.10070826.\u003c/li\u003e\n\u003cli\u003eRahmad NA, Sufri NAJ, Muzamil NH, As\u0026rsquo;ari MA. Badminton player detection using faster region convolutional neural network. Indonesian Journal of Electrical Engineering and Computer Science 2019;14:1330\u0026ndash;5. https://doi.org/10.11591/ijeecs.v14.i3.pp1330-1335.\u003c/li\u003e\n\u003cli\u003eBaclig MM, Ergezinger N, Mei Q, Gul M, Adeeb S, Westover L. A deep learning and computer vision based multi‑player tracker for squash. Applied Sciences 2020;10. https://doi.org/10.3390/app10248793.\u003c/li\u003e\n\u003cli\u003eLiu P, Wang J-H. MonoTrack: Shuttle trajectory reconstruction from monocular badminton video. Proceedings of the IEEE Computer Vision and Pattern Recognition Workshops 2022:3512\u0026ndash;21. https://doi.org/10.1109/CVPRW56347.2022.00395.\u003c/li\u003e\n\u003cli\u003eWu F, Wang Q, Bian J, Ding N, Lu F, Cheng J, et al. A survey on video action recognition in sports: Datasets, methods and applications. IEEE Trans Multimedia 2023;25:7943\u0026ndash;66. https://doi.org/10.1109/TMM.2022.3232034.\u003c/li\u003e\n\u003cli\u003eWei CT, Weng SK. A court line extraction algorithm for badminton tournament videos with horizontal line projection learning. IET Image Process 2023;17:2907\u0026ndash;24. https://doi.org/10.1049/ipr2.12838.\u003c/li\u003e\n\u003cli\u003eZhang L, Dai H. Motion trajectory tracking of athletes with improved depth information‑based KCF tracking method. Multimed Tools Appl 2023;82:26481\u0026ndash;93. https://doi.org/10.1007/s11042-023-14929-6.\u003c/li\u003e\n\u003cli\u003eZhi J, Sun Z, Zhang R, Zhao Z. Badminton video action recognition based on time network. Journal of Computational Methods in Sciences and Engineering 2023;23:2739\u0026ndash;52. https://doi.org/10.3233/JCM-226889.\u003c/li\u003e\n\u003cli\u003eHsu YH, Yu CC, Cheng HY. Enhancing badminton game analysis: An approach to shot refinement via a fusion of shuttlecock tracking and hit detection from monocular camera. Sensors 2024;24. https://doi.org/10.3390/s24134372.\u003c/li\u003e\n\u003cli\u003eAmudhan AN, Vrajesh SR, Sudheer AP, Lijiya A. RFSOD: A lightweight single‑stage detector for real‑time embedded applications to detect small‑size objects. J Real Time Image Process 2022;19:133\u0026ndash;46. https://doi.org/10.1007/s11554-021-01170-3.\u003c/li\u003e\n\u003cli\u003eTan YE, See J, Abdullah J, Wang LK, Ban KW. A deep learning based framework for badminton rally outcome prediction. Proceedings of the International Symposium on Intelligent Signal Processing and Communication Systems 2022. https://doi.org/10.1109/ISPACS57703.2022.10082764.\u003c/li\u003e\n\u003cli\u003eGonzalez MR, Sandoval JC, Tostado GF, Santamaria SS, San Blas HS, Gonzalez G V, et al. Development of a stroke detection system for racket sports. Proceedings of the International Conference on Disruptive Technologies, Tech Ethics, and Artificial Intelligence 2023:320\u0026ndash;8. https://doi.org/10.1007/978-3-031-38344-1_31.\u003c/li\u003e\n\u003cli\u003eXie Y, Xie X, Xia H, Zhao Z. Algorithmic study on position and movement method of badminton doubles. Sci Program 2022. https://doi.org/10.1155/2022/6422442.\u003c/li\u003e\n\u003cli\u003eZhang L, Zhong P. Feather teaching and training based on Kinect sensor and gesture recognition technology. Systems and Soft Computing 2025;7:200349. https://doi.org/10.1016/j.sasc.2025.200349.\u003c/li\u003e\n\u003cli\u003ePeng W, Zheng W. An AI-based badminton smash detection using residual-shuffle network optimized based on upgraded pufferfish optimizer. Ain Shams Engineering Journal 2025;16:103414. https://doi.org/10.1016/j.asej.2025.103414.\u003c/li\u003e\n\u003cli\u003eJiang K. Optimization study of badminton sports training system based on MoileNet OpenPose lightweight human posture estimation model. Entertain Comput 2025;54:100975. https://doi.org/10.1016/j.entcom.2025.100975.\u003c/li\u003e\n\u003cli\u003eZhang J, Han D, Han S, Li H, Lam W-K, Zhang M. ChatMatch: Exploring the potential of hybrid vision\u0026ndash;language deep learning approach for the intelligent analysis and inference of racket sports. Comput Speech Lang 2025;89:101694. https://doi.org/10.1016/j.csl.2024.101694.\u003c/li\u003e\n\u003cli\u003eLi J, Fan Y, Chen R, Liang S, Feng Y, He Y, et al. An Intelligent Badminton Handle With Multinode MEMS Sensors for Explainable Motion Recognition. IEEE Internet Things J 2025;12:37022\u0026ndash;34. https://doi.org/10.1109/JIOT.2025.3580545.\u003c/li\u003e\n\u003cli\u003eShih C-Y, Lin Y-T, Chen W, Huang J-C. SVM-Adaboost based badminton offensive movement parsing technique. Signal Image Video Process 2025;19:280. https://doi.org/10.1007/s11760-025-03865-7.\u003c/li\u003e\n\u003cli\u003eLiu D, Che L, Qi F, Zhang K, Cao D. Evaluating the sports performance of badminton players based on grip strength of the real hitting scenario. Sci Rep 2026;16:5055. https://doi.org/10.1038/s41598-026-35596-1.\u003c/li\u003e\n\u003cli\u003eChen E, Zheng Y. Harnessing deep learning and advanced analytics to revolutionise badminton performance. International Journal of Information and Communication Technology 2025;26:61\u0026ndash;79. https://doi.org/10.1504/IJICT.2025.146097.\u003c/li\u003e\n\u003cli\u003eLiu M. Badminton Technical Training Analysis System Based on Intelligent Motion Tracking Decomposition Model. 2025 International Conference on Algorithm, Artificial Intelligence and Computer Vision (AAICV), IEEE; 2025, p. 01\u0026ndash;4. https://doi.org/10.1109/AAICV66571.2025.00013.\u003c/li\u003e\n\u003cli\u003eTan and J. Komar, \u0026ldquo;How Data Can Help You Beat Your Next Opponent: Applications of Artificial Intelligence for Performance Analysis in Badminton,\u0026rdquo; in Sports Data Analytics: Techniques, Applications, and Innovations, M. S. Raval, T. Kaya, N. S. Artan, and C. Taber, Eds. Singapore: Springer Nature Singapore, 2026, pp. 141\u0026ndash;162. https://doi.org/10.1007/978-981-95-5132-3_8.\u003c/li\u003e\n\u003cli\u003eSingha KAK, Ali MA, Acharjee T, Debnath P, Ahmadullah S. Badminton Match Analysis Along With Players and Shuttlecock Setection Using Hybrid CNN. 2025 IEEE Silchar Subsection Conference (SILCON), IEEE; 2025, p. 1\u0026ndash;6. https://doi.org/10.1109/SILCON67893.2025.11327275.\u003c/li\u003e\n\u003cli\u003eLuo L, Zhang L, Liu H. An integrated deep learning framework for real-time badminton action recognition and pose estimation with shot-level semantic refinement. Australian Journal of Electrical and Electronics Engineering 2025:1\u0026ndash;20. https://doi.org/10.1080/1448837X.2025.2553948.\u003c/li\u003e\n\u003cli\u003eZhu X, Liu L, Huang J, Chen G, Ling X, Chen Y. The analysis of motion recognition model for badminton player movements using machine learning. Sci Rep 2025;15:19030. https://doi.org/10.1038/s41598-025-02771-9.\u003c/li\u003e\n\u003cli\u003eMa H, Zhang F, Liang N. Development and Training Strategy of Badminton Action Recognition System Under the Background of Artificial Intelligence. IEEE Access 2025;13:71577\u0026ndash;85. https://doi.org/10.1109/ACCESS.2025.3563302.\u003c/li\u003e\n\u003cli\u003eChang Y. A method of badminton video motion recognition based on adaptive enhanced AdaBoost algorithm. Int J Biom 2025;17:86\u0026ndash;101. https://doi.org/10.1504/IJBM.2025.143715.\u003c/li\u003e\n\u003cli\u003eHsu J-H, Lee C-C, Chang J-Y, Lee D-S. Key Frame Detection in Badminton Swings and Its Application to Physical Education. IEEE Access 2025;13:91248\u0026ndash;62. https://doi.org/10.1109/ACCESS.2025.3572105.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Artificial intelligence, badminton, computer vision, player assessment, sports analytics","lastPublishedDoi":"10.21203/rs.3.rs-9105158/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9105158/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eArtificial intelligence (AI) is increasingly promoted as a means of providing objective player performance assessment in badminton. Compared to other sports, however, the supporting evidence remains fragmented. A systematic review based on 51 studies that satisfied the established quality and eligibility criteria from three major databases (covering the period 2018 to the end of 2025) reveals four dominating methodological schools: computer vision stroke tracking, movement-pattern recognition, spatio-temporal analysis of rally sequences, and multi-modal frameworks that integrate several data streams. Although many studies report high classification or prediction accuracy, only a small proportion of them employ shared validation datasets or evaluate repeatability across testing sessions, which limits the generalisability of their findings. Common shortcomings include small or imbalanced data samples, weak alignment with established sport-science theory, substantial computational requirements, and participant pools drawn largely from elite athletes in a single geographic region. Recent work has begun integrating explainable AI with retrieval-augmented generation (RAG) and large language model (LLM) frameworks to provide grounded, query-responsive feedback that links visual detections and performance metrics to structured match evidence. Future research should focus on larger and more diverse datasets, alignment with skill development models, transparent output formats, and validation across competitive levels and contexts with these state-of-the-art explainable AI-based RAG or LLM frameworks.\u003c/p\u003e","manuscriptTitle":"Making the World’s Fastest Racket Sport even Better: A Systematic Review of Artificial Intelligence-based Objective Player Performance Assessment in Badminton","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-18 08:46:25","doi":"10.21203/rs.3.rs-9105158/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"57d9fd55-c14d-473b-b831-eeaf5aba461a","owner":[],"postedDate":"March 18th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-18T06:40:16+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-18 08:46:25","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9105158","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9105158","identity":"rs-9105158","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}