Research on intelligent generation of volleyball training strategies combining YOLOv5 + DeepSORT trajectory data and a key point CNN model Authors: Liang Zhao Published: 2026 ABSTRACT Volleyball training increasingly requires objective information about how players move, coordinate, and execute technical actions during live practice. This study evaluated an integrated video-based system combining YOLOv5 player detection, DeepSORT trajectory tracking, and convolutional neural-network keypoint analysis to support individualized training prescriptions. Fifty-six competitive volleyball athletes were assigned to a six-week parallel-group intervention comparing standard coaching with coaching supported by real-time tactical and biomechanical feedback. The system translated player movement, court positioning, jump mechanics, approach angles, arm-swing timing, and block timing into drill recommendations during training sessions. Compared with standard coaching, the video-supported condition showed greater improvements in transition speed, serve-receive accuracy, rotation latency, out-of-system play reduction, defensive coverage, jump height, take-off time, approach mechanics, arm-swing sequencing, and block timing consistency. Scrimmage indicators, including side-out rate, point differential, and attack efficiency, also improved more in the video-supported group, and several tactical and biomechanical gains were better retained after one to two weeks. The findings indicate that combining real-time trajectory and keypoint information with structured drill selection can support more precise volleyball training decisions. The approach may help coaches identify performance deficits, individualize practice content, and monitor short-term learning, while further validation across venues, hardware conditions, and competition settings remains necessary.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-56040-4. Introduction Volleyball is a dynamic, high-intensity sport in which tactical positioning, the coordinated arrangement of players in response to offensive or defensive scenarios, plays a critical role in match outcomes1. Alongside this, team synchronization, the precise temporal alignment of player movements, ensures seamless execution of plays such as quick attacks and coordinated blocks2. To optimize these elements, modern coaching increasingly requires granular insights, meaning fine-grained performance metrics that quantify both technical and tactical proficiency3. Traditional coaching approaches, which often depend on manual video review, are time-consuming and prone to inconsistency due to subjective interpretation. Recent comparative analyses indicate that AI-driven methods in sports like soccer and basketball outnumber volleyball-focused systems by more than 10:1, confirming that volleyball remains comparatively underexplored in AI research4. Bibliometric analyses estimate that volleyball-specific AI studies account for under 7% of all sports computer vision publications in the past five years5. This gap is important because volleyball performance is shaped by interactions between collective court organization and individual technical execution. A serve–receive sequence, transition phase, block formation, and defensive recovery pattern each depend on rapid player repositioning, but the effectiveness of these movements also depends on biomechanical details such as approach angle, jump timing, arm-swing coordination, and landing control. Artificial intelligence (AI) and machine learning (ML) methods can therefore contribute most directly to volleyball training when visual information is converted into interpretable indicators that coaches can use during practice. Existing AI applications in sport have demonstrated strong potential for automated detection and classification, but volleyball requires a system that links court-level movement patterns with athlete-level execution quality in a form that can support drill selection and performance monitoring. ML has emerged as a powerful tool for extracting and analyzing large volumes of multimodal sports data6. In volleyball, ML applications have addressed isolated tasks such as ball detection, attack classification, or defensive formation recognition7. However, the lack of integrated end-to-end systems limits their practical impact on training design8. You Only Look Once (YOLO) models are widely used in real-time object detection due to their high precision and low inference latency9. In this work, YOLOv5 is adopted over newer versions like YOLOv8 to balance speed, computational efficiency, and model robustness under hardware constraints10. Prior comparisons have shown YOLOv5 to deliver competitive accuracy while maintaining faster processing rates on embedded GPU systems, making it well-suited for in-situ sports analytics6. The novelty of the present study does not rest on YOLOv5, Deep Simple Online and Realtime Tracking (DeepSORT), or convolutional neural network (CNN) keypoint estimation as separate techniques, because each component has already been established in computer vision research. Instead, the contribution lies in combining these established components into a volleyball-specific workflow that detects players, maintains player identities, estimates movement-related keypoints, derives tactical and biomechanical indicators, and links those indicators to structured training prescriptions under real-time practice conditions. Player tracking extends detection outputs by maintaining consistent player identities across frames, enabling longitudinal analysis of movements11. DeepSORT, in particular, uses appearance descriptors and motion modeling to produce accurate, continuous trajectories12. These trajectories can be converted into volleyball-specific tactical indicators, including serve–receive formation shifts, defensive rotation patterns, and offensive transition timing13. Advanced spatiotemporal models such as TrajGPT have demonstrated the ability to reconstruct trajectories with 26× higher temporal accuracy and 98% spatial fidelity in sparse data settings14. While similar architectures have been applied to continuous-motion sports, volleyball’s rally-based structure requires segmentation-aware processing to isolate and analyze discrete phases of play15. This adaptation of spatiotemporal modeling to rally-structured data represents a methodological innovation in this study. For volleyball, the practical value of trajectory analysis depends on whether continuous player tracks can be interpreted within the tactical phases of play. The same movement path may have different meanings during serve–receive, defensive transition, counterattack preparation, or block coverage. DeepSORT-based identity preservation, therefore, provides a foundation for tracking player movement over time, but trajectory data must still be transformed into volleyball-specific indicators such as rotation-shift latency, transition timing, court-coverage balance, and out-of-system recovery. This phase-sensitive interpretation helps move tracking outputs beyond descriptive visualization and toward decision support for training. CNNs provide the capability to extract biomechanical features from video through keypoint estimation1. In volleyball, these features include spike approach angles, block timing, jump height, and arm swing kinematics2. Such measurements can be validated against expert-labeled ground truth to ensure biomechanical accuracy. While CNN architectures have also been used for action classification in sports, this work focuses on their use in pose estimation for detailed kinematic profiling4. Previous research has shown that multi-stream and attention-augmented CNNs can improve action recognition accuracy by capturing fine-grained spatial-temporal dependencies. However, few volleyball studies have integrated these outputs into strategy recommendation systems that directly inform training interventions. Biomechanical keypoint information adds a second layer of interpretation because tactical success depends partly on the quality of technical execution. A player may reach the correct tactical location but still lose efficiency through delayed take-off, unstable approach mechanics, excessive arm-swing lag, or inconsistent block timing. CNN-based keypoint estimation can therefore complement trajectory tracking by identifying execution features that are difficult to quantify consistently through manual observation alone. When these features are combined with tactical indicators, the resulting system can support more individualized prescriptions, such as transition drills for delayed court recovery, approach-lane correction for suboptimal attack mechanics, or timing drills for inconsistent blocking actions. Problem statement A clear gap remains in volleyball training analytics because existing systems tend to operate either as descriptive video-analysis tools or as isolated recognition models, rather than as integrated systems that connect player detection, identity-preserved tracking, biomechanical keypoint analysis, and drill prescription. This limits the translation of visual analytics into training decisions, particularly when coaches need rapid information about both tactical coordination and movement execution during live practice. Real-time application also depends on computational throughput, latency control, camera synchronization, and hardware availability; therefore, the present system was designed for graphics processing unit (GPU)-equipped training environments and evaluated under a target operating range of at least 30 frames per second (FPS) and no more than 80 ms end-to-end latency. This ensures that feedback can be integrated directly into live training sessions. There is currently no volleyball-specific, real-time framework that integrates YOLOv5 detection, DeepSORT tracking, and CNN-based keypoint estimation into a unified system capable of producing validated, data-driven, and individualized training strategies. Objectives, research question, and hypotheses The study aimed to: (i) develop and evaluate an integrated volleyball video-analytics workflow combining YOLOv5 detection, DeepSORT tracking, and CNN-based keypoint estimation to extract volleyball-specific tactical and biomechanical indicators; and (ii) use these indicators to generate player-specific training prescriptions defined by pre-established movement, timing, and performance thresholds under real-time practice conditions. The study asks the following research question: How can an integrated YOLOv5–DeepSORT–CNN keypoint pipeline be used to generate accurate, real-time, and player-specific volleyball training strategies from combined trajectory and biomechanical data? Based on these objectives and the RQ, the study hypothesizes that H1: The integrated YOLOv5–DeepSORT–CNN pipeline will achieve higher detection, tracking, and pose estimation accuracy for volleyball-specific movements than existing single-component systems. H2: Training strategies generated from combined trajectory–pose data will produce measurable improvements in targeted player performance metrics during controlled training sessions. Literature review Integrated multi-stage visual analytics for volleyball-specific tactical pattern recognition Recent volleyball analytics have explored integrating detection, tracking, and recognition, but implementations vary widely in scope and depth16. Jiang et al.,1 combined YOLOv7 detection, a memory transformer for trajectory continuity, and transformer-based action classification, achieving MOTA 94.8, IDF1 72.6, and HOTA 73.7 on SportsMOT, plus 90.7% group and 85.5% individual action mAP on a Volleyball dataset. While robust under occlusion, the absence of biomechanical keypoint integration means tactical insights are not linked to execution quality. Li and Zhou17, used ITS-DCRNN on motion-captured joint angles, velocities, and accelerations, reaching 98.58% accuracy, but without an automated detection stage, limiting applicability to real-game footage where body occlusion is frequent. Luo et al.,18 pruned YOLOv5 and optimized DeepSORT for real-time traffic analysis, boosting FPS from 16 to 31 while maintaining 95.3% detection accuracy; however, its fixed-zone assumption in traffic flow limits transfer to volleyball, where player zones are dynamic. Zhao19, fused DWT and STFT features with a parallel CNN and Random Forest to reach 99.61% precision in sensor-based sport recognition, but reliance on wearables limits scalability in broadcast or training video contexts. Volleyball analytics now includes stronger detection, tracking, recognition, and biomechanical-classification methods, with reported outputs such as mean average precision (mAP), multiple object tracking accuracy (MOTA), identity F1 score (IDF1), and higher order tracking accuracy (HOTA). The next research priority is to integrate these accuracy metrics with coach-facing training decisions. A real-time workflow can convert detection reliability, trajectory continuity, and movement-recognition outputs into interpretable tactical and biomechanical feedback that supports volleyball practice. Other domain-transferrable methods show similar constraints. Zhang20, achieved 98.4% accuracy in rope skipping recognition with a CNN–LSTM–DQN pipeline, demonstrating tactical optimization in repetitive movement, but without multi-agent, irregular interactions found in volleyball. Zong21, improved YOLOv5’s mAP to 97.7% with Shuffle Attention and DCNv3, enhancing small-object detection useful for distant court views, but omitting pose estimation. Suartika22, optimized YOLOv5 for Raspberry Pi, enabling on-device inference in real time, yet lacked multi-player tracking or biomechanics. Across these works, the main limitation is that while accuracy, speed, and tracking robustness improve individually, no system integrates detection, dynamic trajectory mapping, and biomechanical pose analysis into a feedback-driven tactical framework specific to volleyball. Automated generation of individualized volleyball training strategies from combined trajectory–pose data The translation of trajectory and pose data into adaptive training strategies remains rare. Li et al.,23, using mixed methods with 150 university athletes, identified motivation, coach–athlete relationship, and team cohesion as key performance predictors (R² = 0.71), yet their framework lacked integration with biomechanical or tactical datasets, limiting real-time appliThe present study is positioned as a volleyball-specific integration framework built from established computer vision components. YOLOv5 provides player detection, DeepSORT supports identity-preserved movement trajectories, and CNN-based keypoint estimation extracts execution-related movement features. The combined pipeline links detections to trajectories, trajectories to tactical indicators, keypoints to biomechanical indicators, and both indicator streams to rule-based drill selection. Novelty is therefore located in the applied arrangement, real-time implementation, and training translation of these components within volleyball practice. Loureiro24, ran a randomized trial with U13 female players, comparing daily-adjusted plans to rigid 4-week schedules using GPAI-coded video review. The adaptive group improved GPAI scores 15–20% faster, proving the value of immediate corrections, but these were manually observed, creating scalability and latency issues. Rada et al.,25 showed that year-long plyometric interventions improved vertical jump height by 7–25% using Kistler Force Platform data, supporting biomechanical tailoring; however, it addressed only strength-speed qualities without linking to tactical decision-making. Adaptive planning and biomechanical conditioning both provide useful foundations for individualized volleyball training. Manual Game Performance Assessment Instrument (GPAI)-based review supports flexible tactical planning, while force-platform testing quantifies strength–speed adaptations. A combined trajectory–pose system can link these two forms of evidence inside the same practice environment by recording how athletes move through tactical phases and how effectively they execute technical actions. Such integration supports drill selection based on court behavior, timing quality, movement mechanics, and player-specific performance thresholds. From a predictive standpoint, Sanjaykumar et al.,26 modeled performance in female players using anthropometric and composition data, with Random Forest reaching 94.18% accuracy. While physical profiling can complement pose–trajectory inputs, the models were trained on static metrics, lacking responsiveness to in-game performance fluctuations. Collectively, these studies show that although tactical flexibility, biomechanical conditioning, and physical profiling each enhance performance, their isolated use prevents unified, real-time feedback loops. The defining limitation across prior work is the absence of automated, scalable systems that adapt training prescriptions in real time by combining spatial-temporal patterns, biomechanical indicators, and player-specific readiness. Individualized volleyball training benefits from a structured pathway in which visual data are converted into actionable indicators, indicators are compared with predefined tactical and kinematic thresholds, and threshold deviations trigger corrective drills. The pathway supports coaching decisions by providing standardized measurement, timely prompts, and session-level performance summaries. Coach expertise remains central to contextual interpretation, athlete communication, and management of workload, fatigue, and technical progression. Conceptual model The present study is positioned as a volleyball-specific integration framework built from established computer vision components. YOLOv5 provides player detection, DeepSORT supports identity-preserved movement trajectories, and CNN-based keypoint estimation extracts execution-related movement features. The combined pipeline links detections to trajectories, trajectories to tactical indicators, keypoints to biomechanical indicators, and both indicator streams to rule-based drill selection. Novelty is therefore located in the applied arrangement, real-time implementation, and training translation of these components within volleyball practice. The proposed model directly addresses gaps from previous studies by integrating three modules (see Fig. 1): YOLOv5 detection to resolve the lack of efficient, accurate player identification; DeepSORT trajectory mapping to overcome the absence of dynamic movement context; and CNN-based keypoint analysis to fill the biomechanical execution gap. These feed into a strategy generation layer that bridges the feedback void, converting combined metrics into real-time, individualized training recommendations. By linking tactical recognition with biomechanical quality in a single pipeline, the model delivers what previous systems lacked: a volleyball-specific, adaptive, feedback-driven framework capable of operating at ≥ 30 FPS in live training settings. The conceptual model functions as an applied integration framework. The detection module identifies athletes, the tracking module preserves player identity across rally phases, the keypoint module estimates movement features relevant to volleyball execution, and the strategy-generation module converts combined outputs into drill recommendations. The model connects visual analytics with adaptive training by linking player detection, tactical movement interpretation, biomechanical profiling, and real-time prescription within one volleyball-specific workflow. Fig. 1Conceptual model of the integrated volleyball video-analytics pipeline showing how 4 K/60 fps video streams feed YOLOv5 detection, DeepSORT tracking, CNN keypoint extraction, synchronization fusion, tactical and biomechanical metric generation, and a strategy engine that updates coach dashboards and data logs. Conceptual model of the integrated volleyball video-analytics pipeline showing how 4 K/60 fps video streams feed YOLOv5 detection, DeepSORT tracking, CNN keypoint extraction, synchronization fusion, tactical and biomechanical metric generation, and a strategy engine that updates coach dashboards and data logs. Methods Study design This was a parallel-group, assessor-blinded randomized controlled trial conducted on an indoor volleyball court over six weeks (three sessions/week). Teams were allocated to an AI-assisted intervention or a standard-coaching control, with identical drills, volumes, and coach–athlete ratios. Four synchronized cameras captured 4 K at 60 fps; video streams were analyzed in real time by an integrated pipeline YOLOv5 for detection, DeepSORT for multi-object tracking, and a CNN keypoint model for biomechanics with latency targeted ≤ 80 ms at ≥ 30 fps. The design consisted of (i) pre-intervention baselines for tactical (serve–receive stability, transition time) and biomechanical metrics (jump height, approach angle), (ii) continuous in-session analytics to trigger drill prescriptions in the intervention arm, (iii) matched training exposures across arms, (iv) post-intervention re-assessment under the same protocol, and (v) independent assessor annotations on a fixed sample of frames for reference labels. Primary outcomes were changes in tactical execution and biomechanical efficiency; secondary outcomes were detection/tracking/pose accuracy. All training prescriptions were recorded at the drill, trigger, timing, and progression level. Each record contained the triggering metric, threshold crossed, rally number, player identifier, drill selected, repetitions, sets, rest interval, coach cue category, prescription completion status, and progression or regression decision. Standard coaching sessions matched the intervention arm for session duration, drill category, rest structure, coach–athlete ratio, and cue frequency. Exposure logs separated automated metric detection, dashboard prescription, coach-delivered cueing, drill completion, and progression or regression decisions for each session. This allowed intervention delivery to be quantified by component rather than treated as a single undifferentiated exposure. Ethics The study was approved by the Institutional Review Board of Henan Institute of Physical Education before recruitment; all athletes provided written informed consent, data were anonymized using coded IDs, storage was on encrypted servers with access controls, participation was voluntary with the right to withdraw, and procedures complied with the Declaration of Helsinki. Written informed consent for publication of participant images and any potentially identifying visual information included in Figure S1 was obtained from the participants and, where applicable, their legal guardians. Participants Competitive indoor volleyball athletes were recruited from university and club teams in the same league season; training schedules and facilities were comparable across sites, and sessions were held on regulation courts under standard lighting (see Figure S1A–C). Recruitment targeted starters and regular substitutes to ensure sufficient rally volume per session, with goal balance across positions (setters, middles, outsides, opposites, liberos). Screening and pre-enrolment logs were managed in REDCap (Vanderbilt University, USA) to capture demographics, training age, injury history, and availability. Recruitment Athletes were approached through athletic directors and head coaches by email and in-person briefings after practice; a study flyer with a QR code linked to the REDCap screening form (Vanderbilt University, USA). Interested players received a 10-minute on-site briefing, a schedule outline, and a short video showing the camera layout and dashboard. Eligible volunteers were contacted within 48 h to confirm availability for the six-week window and to arrange baseline recording slots. An a priori sample size target of 56 athletes (28 per arm) was set to detect the expected between-arm differences in primary outcomes, with a 15% attrition allowance yielding a minimum analyzable cohort of 48 athletes. Eligibility criteria Eligibility was verified as follows: inclusion were (i) active roster status via coach-signed official list; (ii) age 18–30 via government ID; (iii) ≥ 2 years formal training via coach log; (iv) safety clearance for high-intensity volleyball participation based on the Physical Activity Readiness Questionnaire Plus (PAR-Q+) with no unresolved positive items, absence of contraindicating symptoms during a standardized 12 min volleyball warm-up, and Borg Category Ratio 10 (CR10) exertion score ≤ 5 recorded within 3 min after warm-up completion. Athletes with positive PAR-Q+ responses, dizziness, chest discomfort, abnormal breathlessness, current pain ≥ 3/10, or Borg CR10 > 5 after the warm-up were deferred until medical clearance or symptom resolution (Warburton et al., 2011; Borg, 1998); (v) written consent to continuous video recording. Exclusion were (i) musculoskeletal injury or concussion within 8 weeks, screened with the OSTRC Overuse Injury Questionnaire (severity score ≥ 20 triggers exclusion) and SCAT6 symptom checklist unless physician clearance27,28; (ii) current pain ≥ 3/10 on activity using the NRS-11 post-warm-up29; (iii) cardiopulmonary/neurological/metabolic conditions prompting PAR-Q+/ePARmed-X+ referral and medical clearance before participation30; (iv) concurrent interventional study participation by self-report/coach verification; (v) projected attendance < 80% based on scheduling commitment. Stratification and bias Random allocation used computer-generated permuted blocks (block sizes of 4–8) with position-stratification to balance setters, middles, outsides/opposites, and liberos across arms; assignments were concealed in sequentially numbered, opaque, tamper-evident envelopes opened by an administrator not involved in assessments. Outcome analysts’ labeling frames and computing metrics were blinded to group identity via anonymized file names; coaches delivering drills could not be blinded to the intervention but were restricted from the AI dashboard in the control arm, and all coach feedback in both arms followed a fixed script to minimize expectancy bias. Coach expectancy was controlled through a standardized 45 min orientation delivered separately for intervention and control coaches. Both groups received identical study information regarding training duration, drill categories, athlete safety rules, and outcome-assessment timing. Intervention coaches were trained only to follow dashboard-generated drill tiles and were not permitted to alter trigger thresholds, algorithm weights, or progression rules. Control coaches delivered the same drill families, session duration, cue frequency, rest intervals, and coach–athlete contact time using a printed session plan without dashboard access. Coaching fidelity was assessed in 20% of randomly selected sessions per arm using video and audio review. The fidelity checklist recorded prescribed drill completion, cue category, cue frequency, rest duration, feedback timing, and deviations from the session plan. Sessions with fidelity < 90% triggered retraining before the next session, and sensitivity analyses excluded sessions below this threshold. System implementation Perception stack Model development used 38,400 annotated volleyball frames sampled from serve–receive, transition, attack, block, and defensive coverage phases. Frames were sampled at 1 frame per second from 32 recorded practice sessions and balanced across camera angle, player position, rally phase, lighting condition, and occlusion level. Annotation was completed in Computer Vision Annotation Tool version 2.8 using bounding boxes for players and ball location, track identity labels, and 17 body keypoints following the Common Objects in Context 17-point convention: nose, left and right eyes, left and right ears, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles. Two trained annotators independently labeled 20% of the frames, and disagreements exceeding 5 pixels for keypoints or 0.10 intersection-over-union units for bounding boxes were reviewed by a senior assessor. Data were split at the athlete and session levels to prevent leakage, with 26,880 frames assigned to training, 5,760 to validation, and 5,760 to testing. No athlete, session, or rally segment appeared in more than one split. Inference ran on Ubuntu 22.04/Python 3.10 with PyTorch 2.2, CUDA 12.1, OpenCV 4.8 on an RTX 4090. YOLOv5 v6.2 was trained at 1280 × 736 with mosaic/HSV augments and AutoAnchor; loss combined DIoU and BCE: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{L}={\lambda\:}_{box}\left(1-\mathrm{I}\mathrm{o}\mathrm{U}+{\rho\:}^{2}/{c}^{2}\right)+{\lambda\:}_{obj}\mathrm{B}\mathrm{C}\mathrm{E}+{\lambda\:}_{cls}\mathrm{B}\mathrm{C}\mathrm{E}$$\end{document}. DeepSORT used a MobileNetV2 cosine-embedder with a constant-velocity Kalman filter; association cost blended motion and appearance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{ij}=\alpha\:\hspace{0.17em}{d}_{ij}^{mah}+\left(1-\alpha\:\right)\left(1-\mathrm{c}\mathrm{o}\mathrm{s}\left({\phi\:}_{i},{\phi\:}_{j}\right)\right)$$\end{document}. A residual encoder–decoder pose CNN estimated 17 joints at 256 × 256 with keypoint loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{kp}=\frac{1}{K}\sum\:\parallel\:{\widehat{\mathbf{y}}}_{k}-{\mathbf{y}}_{k}{\parallel\:}_{2}^{2}+{\lambda\:}_{b}\sum\:_{\left(i,j\right)}{\left(\parallel\:{\widehat{\mathbf{y}}}_{i}-{\widehat{\mathbf{y}}}_{j}{\parallel\:}_{2}-{\mathcal{l}}_{ij}\right)}^{2}$$\end{document}; kinematics (jump height, approach angle, arm-swing lag) were derived per track. YOLOv5 v6.2 training used stochastic gradient descent with momentum 0.937, weight decay 0.0005, initial learning rate 0.01, cosine learning-rate decay, batch size 16, image size 1280 × 736, 120 epochs, mosaic augmentation probability 0.80, hue-saturation-value augmentation probability 0.50, horizontal flip probability 0.50, and early stopping after 20 epochs without validation improvement. DeepSORT appearance embeddings were fine-tuned with triplet loss using margin 0.30, Adam optimizer, learning rate 0.0003, batch size 64, and 60 epochs. The pose CNN was trained at 256 × 256 resolution using AdamW, learning rate 0.0001, weight decay 0.0001, batch size 32, 100 epochs, Gaussian heatmaps with σ = 2 pixels, and keypoint visibility masking for occluded joints. Model checkpoints were selected using validation mean average precision for detection, identity F1 score for tracking, and percentage of correct keypoints at 0.2 torso length for pose estimation. Algorithm 1: Perception Stack. Algorithm 1: Perception Stack. Computer vision performance was evaluated on the athlete-independent test set before intervention analyses. Detection was quantified using mean average precision at intersection-over-union thresholds of 0.50 and 0.50:0.95, with separate reporting for player and ball classes. Tracking was quantified using multiple object tracking accuracy, identity F1 score, higher order tracking accuracy, identity switches per 1,000 frames, and track fragmentation rate. Keypoint estimation was quantified using percentage of correct keypoints at 0.2 torso length, normalized mean error, and joint-angle mean absolute error for shoulder, elbow, hip, knee, and ankle angles. Confidence intervals were estimated with 1,000 bootstrap resamples at the rally level. Inter-rater reliability for assessor-derived reference labels was evaluated on the double-annotated 20% frame subset; final agreement reached ICC = 0.94 for continuous keypoint coordinates, ICC = 0.89 for joint-angle measures, Cohen’s κ = 0.87 for categorical rally-phase labels, and mean bounding-box intersection-over-union agreement = 0.92, exceeding the predefined acceptance thresholds of ICC ≥ 0.85, κ ≥ 0.80, and mean intersection-over-union ≥ 0.85. Calibration, synchronization, and performance benchmarks Cameras were calibrated with the Zhang checkerboard via calibrateCamera, correcting radial/tangential distortion and mapping to court homographies; extrinsics were rechecked weekly. Pixel coordinates were converted to court-referenced metric coordinates using a planar homography estimated from the four court corners, centerline intersections, attack-line intersections, and service-zone boundary points on the 18 m × 9 m regulation court. A 7 × 9 checkerboard with 30 mm squares was recorded at the start of each session to estimate intrinsics, and court-line control points were reprojected after calibration. Sessions were accepted when mean reprojection error was ≤ 0.8 pixels and court-distance reconstruction error was ≤ 2.0 cm across six known line distances. Athlete standing height, measured to the nearest 0.5 cm with a wall stadiometer, was used only as a secondary scale constraint for vertical keypoint normalization. Jump height was derived from frame-wise ankle–hip center displacement and take-off/landing timing, then corrected using court-plane scale and athlete-specific standing calibration frames. Approach angle was calculated from the last two approach steps relative to the attack-line coordinate frame. Arm-swing lag was calculated as the time difference between peak elbow extension velocity and peak wrist forward velocity. Shoulder angular velocity was calculated from the shoulder–elbow vector using smoothed keypoint trajectories filtered with a fourth-order Butterworth low-pass filter at 6 Hz. Streams were synchronized using an LED flash visible to all cameras and NTP-disciplined clocks; offsets were \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta t_{i} = t_{i}^{{LED}} - \bar{t}^{{LED}}$$\end{document} End-to-end latency was measured as the time from camera exposure timestamp to dashboard update timestamp, and module latency was measured with synchronized CUDA events and system clock markers for detection, tracking, pose estimation, kinematic calculation, prescription generation, and dashboard rendering. Benchmarking was performed during three uninterrupted 180 min training sessions and one 30 min stress-test session containing dense rallies, crossing players, and partial occlusions. Latency was summarized as mean, standard deviation, median, 95th percentile, and 99th percentile for each module and for the full pipeline. Throughput was computed as processed frames per second per stream and across all streams. Stability logs recorded dropped-frame rate, graphics processing unit utilization, central processing unit utilization, video memory use, system memory use, thermal throttling events, and synchronization drift. The system was considered real-time when median throughput remained ≥ 30 frames per second, 95th percentile end-to-end latency remained ≤ 80 ms, dropped-frame rate remained ≤ 1%, and synchronization drift remained ≤ 10 ms over a full session. Algorithm 2: Calibration, synchronization, and benchmarking. Algorithm 2: Calibration, synchronization, and benchmarking. A biomechanical validation subset was collected in 12 athletes selected to represent setters, middles, outside hitters, opposites, and liberos. Each athlete completed 10 spike approaches, 10 block jumps, and 10 defensive transition trials while recorded simultaneously by the four-camera video system and an eight-camera optical motion-capture system operating at 200 Hz, with two embedded force plates sampled at 1,000 Hz. Reflective markers were placed bilaterally on the acromion, lateral epicondyle, wrist styloid, anterior superior iliac spine, lateral femoral epicondyle, lateral malleolus, heel, and second metatarsal head. Synchronization used a visible light-emitting diode pulse and a transistor–transistor logic trigger recorded by the motion-capture system. Video-derived jump height, approach angle, arm-swing lag, shoulder angular velocity, and block-timing standard deviation were compared with motion-capture and force-plate references using mean absolute error, root mean square error, Bland–Altman bias, 95% limits of agreement, and intraclass correlation coefficients. A priori validation thresholds were mean absolute error ≤ 3.0 cm for jump height, ≤ 3.0° for approach angle, ≤ 15 ms for timing variables, ≤ 80 °/s for shoulder angular velocity, and intraclass correlation coefficient ≥ 0.75. The downscaling protocol was evaluated before deployment under three operating profiles: 4 K/60 frames per second full resolution, 1920 × 1080/60 frames per second downscaled resolution, and region-of-interest cropping around the court at 1920 × 1080/60 frames per second. Automatic fallback was triggered when 95th percentile latency exceeded 80 ms for two consecutive 60 s windows, dropped-frame rate exceeded 1%, or graphics processing unit memory use exceeded 90%. The fallback sequence first enabled half-precision inference, then activated court-region cropping, and finally reduced input resolution to 1920 × 1080 while preserving 60 fps capture. The selected fallback profile was retained only if tracking identity F1 score decreased by no more than 5% points and keypoint percentage of correct keypoints at 0.2 torso length decreased by no more than 5% points relative to the 4 K profile. Implementation burden was recorded for each session to characterize practical deployment requirements. Setup time was defined as camera placement, tripod leveling, cable routing, synchronization check, and calibration recording, with a target duration ≤ 60 min before the first session and ≤ 20 min for routine weekly recalibration. Minimum staffing required one system operator and one coach per court. Hardware consisted of four 4 K/60 frames per second cameras, four tripods placed outside the playable area, one synchronization light-emitting diode unit, one graphics processing unit workstation, and one courtside tablet or monitor. Maintenance logs recorded camera repositioning, recalibration time, storage volume, software restart events, failed recordings, and operator interventions. A lower-resource profile using 1920 × 1080/60 frames per second input and region-of-interest cropping was evaluated as the fallback configuration for facilities without sustained 4 K processing capacity. Strategy generation and generalization Windowed rally features f (transition time, approach angle, arm-lag, block-timing SD) were standardized per athlete \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:z=\left(f-\mu\:\right)/\sigma\:$$\end{document} and combined \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S = w^{{ \top }} z$$\end{document}; prescriptions triggered when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:S>\theta\:$$\end{document} for r consecutive windows or on hard-rule violations (e.g., approach angle < 18°, transition > 1.20 s). Algorithm 3: Strategy generation and OOD robustness. Algorithm 3: Strategy generation and OOD robustness. Generalization was evaluated using an out-of-distribution test set collected from three additional indoor courts that differed in lighting intensity, floor color, wall background, and camera height. The out-of-distribution set included 9,600 frames from 12 athletes who were not included in model training, validation, testing, or intervention analyses. Recording conditions included two-camera and four-camera layouts, 4 K/60 frames per second and 1920 × 1080/60 frames per second profiles, and athletes from university and club competitive levels. Detection, tracking, and pose metrics were computed using the same procedures as the in-distribution test set. Robustness was considered acceptable when out-of-distribution degradation relative to in-distribution testing remained within 5% points for mean average precision, 7% points for identity F1 score, 7% points for higher order tracking accuracy, and 5% points for percentage of correct keypoints. Sessions exceeding these margins were assigned to error analysis by occlusion type, camera angle, lighting condition, and athlete density. Training intervention protocol Drill integration Deficiencies triggered preset prescriptions. Transition time > 1.20 s (≥ 3 rallies): cone shuffle–crossover–sprint, 3 × 8; ladder first-step, 3 × 30 s. Approach angle < 18° (≥ 3 spikes): three-step lane rebuild (20°/25°), 4 × 6; med-ball approach-to-throw (2–3 kg), 3 × 8. Arm-swing lag > 70 ms: band scapular retraction, 3 × 12; metronome tempo arm-sequencing, 3 × 30 s. Block-timing SD > 90 ms: LED-cued synchronized block jumps, 3 × 10; read-and-react closes, 3 × 8. Prescriptions appeared on the courtside dashboard and were logged automatically. Feedback frequency Real-time feedback was issued after each rally (latency ≤ 80 ms) with color status and metric values; only triggered items produced drill tiles. End-of-session summaries compiled medians/percentiles and a prioritized drill list; reports were archived to the study database. Adaptation mechanism Progression occurred when a target metric improved ≥ 10% from session baseline and remained in-band for two consecutive sets: volume + 2 reps or complexity + 1 level or rest − 15 s. Regression rules activated when performance worsened ≥ 5% or stayed out-of-band for two sets: revert to precursor drill, add rest + 15–30 s, or switch to part-practice. Safety guards overrode progression: reported pain ≥ 3/10 or two missed landings in a set prompted immediate step-down and technique-only work. All changes were time-stamped and versioned. Intervention fidelity monitoring Fidelity was tracked via a per-drill checklist (delivered, reps/sets, cues, rest) auto-verified against timestamps; a fidelity score (% completed/expected) was computed per session, with remediation required if < 90%. Coaches completed a brief calibration module; 10% of drills were audited weekly for cue consistency. Control coaches had no access to dashboard metrics, threshold flags, prescription logs, or end-of-session automated reports. Intervention coaches viewed only athlete identifier, triggered metric category, traffic-light status, and assigned drill tile during dead-ball intervals; continuous raw metric streams, comparative group statistics, and outcome-test results were not displayed during live coaching. Dashboard access logs recorded screen-open time, tile-view duration, prescription acceptance, prescription override, and manual note entry. Weekly audits compared dashboard logs with session video to verify whether prescriptions were delivered as displayed. Coach and athlete usability assessment At the end of week 6, dashboard usability, cognitive load, and perceived training value were assessed among intervention athletes and coaches. Athletes completed a 10 item usability questionnaire adapted from the System Usability Scale using a 5 point response scale, plus four volleyball-specific items on clarity of feedback, perceived relevance of drills, timing of feedback, and training burden. Coaches completed the National Aeronautics and Space Administration Task Load Index after the final intervention week and a 15 min semi-structured interview focused on dashboard readability, drill selection, cognitive load, integration into practice flow, and perceived barriers to implementation. Interviews were audio-recorded, transcribed verbatim, anonymized, and coded by two independent assessors using deductive categories of usability, workload, timing, interpretability, and practical implementation. Coding agreement was quantified using Cohen’s kappa, with κ ≥ 0.80 treated as acceptable before thematic summaries were finalized. Statistical analysis All analyses were conducted in Python 3.10 (Python Software Foundation, USA). Detection performance was summarized by mAP@0.5, mAP@0.5:0.95, and FPS; tracking by MOTA, IDF1, and HOTA; keypoints by PCK@0.2 and joint-angle MAE; training outcomes were computed as per-athlete percent change from baseline for tactical (e.g., transition time) and biomechanical metrics (e.g., approach angle). Between-group effects were tested with mixed-effects ANOVA (fixed factors: group, time; random intercept: subject) and, where appropriate, within-subject pre–post comparisons with paired t-tests; effect sizes (Cohen’s d; partial η²) and 95% CIs were reported. H1 was evaluated by comparing the integrated pipeline’s accuracy metrics against baseline systems; H2 was evaluated by comparing improvement magnitudes between groups. Assumptions (normality, homogeneity) were checked on residuals; if violated, nonparametric alternatives or robust estimators were applied, and sensitivity analyses repeated results excluding sessions with fidelity < 90% and under alternative trigger thresholds. Longitudinal change across sessions was modeled with repeated-measures mixed models to estimate time×group interactions. Primary outcomes were defined before analysis as transition time, serve–receive accuracy, rotation-shift latency, out-of-system rate, defensive coverage index, jump height, approach angle, arm-swing lag, and block-timing standard deviation. Secondary and exploratory outcomes included responder rates, dose–response slopes, retention outcomes, and position-specific subgroup estimates. Multiplicity across primary outcomes was controlled using the Benjamini–Hochberg false discovery rate procedure at q = 0.05, while secondary outcomes were interpreted as exploratory and reported with unadjusted p values, effect sizes, and 95% confidence intervals. Responder thresholds were specified before analysis as ≥ 10% reduction in transition time, ≥ 5%-point increase in serve–receive accuracy, ≥ 0.08 s reduction in rotation-shift latency, ≥ 0.05 increase in defensive coverage index, ≥ 5 cm increase in jump height, and ≥ 15% reduction in block-timing standard deviation. Sensitivity analyses repeated the primary models using Holm-adjusted p values to assess robustness of statistical interpretation. All tests were two-tailed with α = 0.05. Results Baseline characteristics At baseline, the RQ is answered by confirming whether the two groups entered the intervention with comparable tactical, biomechanical, and training-exposure profiles before evaluating the YOLOv5–DeepSORT–CNN training pipeline. H2 is tested in later sections through MDs in pre–post change between groups. MD refers to the between-group difference in change scores, calculated as intervention-group change minus control-group change; negative MDs show larger reductions in time, error, latency, or variability, while positive MDs show larger gains in accuracy, coverage, jump height, or successful performance. Table 1Baseline characteristics, tactical, and biomechanical performance.VariableIntervention (n = 28)Control (n = 28)Sig. Demographics & training Age (years)21.6 ± 2.121.9 ± 2.00.585Training age (years)6.1 ± 2.06.3 ± 1.90.705Weekly sessions4.2 ± 0.84.3 ± 0.70.621Weekly minutes390 ± 60400 ± 550.518Prior-season sets42 ± 1240 ± 130.552 Tactical performance Serve–receive stability (%)72.4 ± 9.171.6 ± 9.40.749Transition time (s)1.31 ± 0.121.33 ± 0.130.552Rotation-shift latency (s)0.84 ± 0.100.86 ± 0.110.480Defensive coverage index (0–1)0.64 ± 0.060.63 ± 0.070.569Out-of-system rate (%)18.5 ± 4.219.1 ± 4.00.587Reception error rate (%)10.2 ± 3.610.6 ± 3.80.688 Biomechanical performance Jump height (cm)49.8 ± 7.049.1 ± 7.20.712Take-off time (ms)310 ± 35315 ± 360.600Approach angle (°)20.1 ± 2.819.8 ± 2.90.696Step length ratio1.35 ± 0.121.33 ± 0.110.518Arm-swing lag (ms)64 ± 1266 ± 130.552Peak shoulder angular vel. (°/s)980 ± 140965 ± 1500.701Block-timing SD (ms)105 ± 18107 ± 190.688Data is presented as mean ± SD. Table 1Baseline characteristics, tactical, and biomechanical performance.VariableIntervention (n = 28)Control (n = 28)Sig. Demographics & training Age (years)21.6 ± 2.121.9 ± 2.00.585Training age (years)6.1 ± 2.06.3 ± 1.90.705Weekly sessions4.2 ± 0.84.3 ± 0.70.621Weekly minutes390 ± 60400 ± 550.518Prior-season sets42 ± 1240 ± 130.552 Tactical performance Serve–receive stability (%)72.4 ± 9.171.6 ± 9.40.749Transition time (s)1.31 ± 0.121.33 ± 0.130.552Rotation-shift latency (s)0.84 ± 0.100.86 ± 0.110.480Defensive coverage index (0–1)0.64 ± 0.060.63 ± 0.070.569Out-of-system rate (%)18.5 ± 4.219.1 ± 4.00.587Reception error rate (%)10.2 ± 3.610.6 ± 3.80.688 Biomechanical performance Jump height (cm)49.8 ± 7.049.1 ± 7.20.712Take-off time (ms)310 ± 35315 ± 360.600Approach angle (°)20.1 ± 2.819.8 ± 2.90.696Step length ratio1.35 ± 0.121.33 ± 0.110.518Arm-swing lag (ms)64 ± 1266 ± 130.552Peak shoulder angular vel. (°/s)980 ± 140965 ± 1500.701Block-timing SD (ms)105 ± 18107 ± 190.688Data is presented as mean ± SD. Baseline characteristics, tactical, and biomechanical performance. Data is presented as mean ± SD. Age showed a small baseline difference favoring slightly younger intervention athletes (SMD = − 0.150, p = 0.585), while training age was similarly balanced between groups (SMD = − 0.102, p = 0.705) (Table 1). Weekly sessions (SMD = − 0.133, p = 0.621) and weekly training minutes (SMD = − 0.174, p = 0.518) showed closely matched exposure before the intervention. Tactical starting points were also balanced, including defensive coverage index (SMD = 0.153, p = 0.569), transition time (SMD = − 0.160, p = 0.552), and out-of-system rate (SMD = − 0.146, p = 0.587). Biomechanical starting points showed similarly small differences for jump height (SMD = 0.099, p = 0.712), arm-swing lag (SMD = − 0.160, p = 0.552), and block-timing SD (SMD = − 0.108, p = 0.688). All absolute SMDs were < 0.20, indicating trivial baseline imbalance across training exposure, tactical execution, and biomechanical performance. Tactical performance In tactical performance, the RQ is answered by examining whether trajectory-derived prescriptions improved transition speed, serve–receive organization, rotation-shift timing, out-of-system recovery, and defensive coverage. H2 is tested through between-group pre–post change analyses for each tactical endpoint, supplemented by threshold-attainment and responder analyses for transition time and out-of-system reduction. Transition time decreased by − 0.13 ± 0.08 s in the intervention group and − 0.05 ± 0.09 s in the control group, with MD = − 0.08 s (95% CI − 0.126, − 0.034; p = 0.003). The MD equalled 6.1% of the intervention group’s baseline transition time, indicating faster court recovery during a rally phase measured in fractions of a second. SMD = − 0.82 indicates a large intervention advantage in transition-speed change. Target attainment ≤ 1.20 s reached 78.6% in the intervention group and 50.0% in the control group, while ≥ 10% transition responders were 67.9% and 32.1%, respectively (p = 0.005), indicating that the intervention group produced approximately twice the responder proportion (see Table 2). Serve–receive accuracy increased by + 8.1 ± 6.0 pp in the intervention group and + 2.4 ± 5.8 pp in the control group, with MD = + 5.7 pp (95% CI + 2.1, + 9.3; p = 0.002). SMD = 0.87 indicates a large intervention advantage in formation correctness. Reception errors decreased by MD = − 1.2 pp (95% CI − 2.1, − 0.3; p = 0.007), and set-level variability decreased by MD = − 5.9 pp (95% CI − 9.3, − 2.5; p = 0.001). The serve–receive results indicate higher formation accuracy, fewer reception errors, and more stable organization across repeated sets. Table 2Primary tactical outcomes.DomainOutcomeInterventionControlMD or RD (95% CI) p EffectTransition speedTransition time (s)−0.13 ± 0.08−0.05 ± 0.09MD = − 0.08 s (− 0.126, − 0.034)0.003SMD = − 0.82Within ≤ 1.20 s, n/N (%)22/28 (78.6%)14/28 (50.0%)RD = 0.286 (0.047, 0.525)0.012RR = 1.57≥ 10% responder rate, n/N (%)19/28 (67.9%)9/28 (32.1%)RD = 0.358 (0.149, 0.567)0.005RR = 2.12Serve–receive stabilityCorrect formation (pp)+ 8.1 ± 6.0+ 2.4 ± 5.8MD = + 5.7 pp (+ 2.1, + 9.3)0.002SMD = 0.87Reception errors (pp)−1.8 ± 2.6−0.6 ± 2.6MD = − 1.2 pp (− 2.1, − 0.3)0.007SMD = − 0.74Consistency CV (pp)−7.0 ± 6.0−1.1 ± 6.0MD = − 5.9 pp (− 9.3, − 2.5)0.001SMD = − 0.93Rotation and system efficiencyRotation-shift latency (s)−0.10 ± 0.06−0.04 ± 0.06MD = − 0.06 s (− 0.092, − 0.028)0.001SMD = − 1.00Out-of-system rate (pp)−4.5 ± 3.8−1.3 ± 3.7MD = − 3.2 pp (− 5.2, − 1.2)0.002SMD = − 0.86≥ 20% OOS reduction, n/N (%)17/28 (60.7%)7/28 (25.0%)RD = 0.357 (0.118, 0.596)0.002RR = 2.43Defensive coverageCoverage index (0–1)+ 0.06 ± 0.05+ 0.02 ± 0.05MD = + 0.04 (+ 0.013, + 0.067)0.004SMD = 0.78Successful digs (pp)+ 7.0 ± 7.0+ 2.0 ± 7.0MD = + 5.0 pp (+ 1.2, + 8.8)0.010SMD = 0.71Coverage uniformity CV (pp)−6.0 ± 5.0−2.0 ± 5.0MD = − 4.0 pp (− 6.7, − 1.3)0.004SMD = − 0.78MD, mean difference in pre–post change between groups, calculated as intervention change minus control change. RD , risk difference. RR , risk ratio. SMD, standardized mean difference. OOS, out-of-system. pp, percentage points. Table 2Primary tactical outcomes.DomainOutcomeInterventionControlMD or RD (95% CI) p EffectTransition speedTransition time (s)−0.13 ± 0.08−0.05 ± 0.09MD = − 0.08 s (− 0.126, − 0.034)0.003SMD = − 0.82Within ≤ 1.20 s, n/N (%)22/28 (78.6%)14/28 (50.0%)RD = 0.286 (0.047, 0.525)0.012RR = 1.57≥ 10% responder rate, n/N (%)19/28 (67.9%)9/28 (32.1%)RD = 0.358 (0.149, 0.567)0.005RR = 2.12Serve–receive stabilityCorrect formation (pp)+ 8.1 ± 6.0+ 2.4 ± 5.8MD = + 5.7 pp (+ 2.1, + 9.3)0.002SMD = 0.87Reception errors (pp)−1.8 ± 2.6−0.6 ± 2.6MD = − 1.2 pp (− 2.1, − 0.3)0.007SMD = − 0.74Consistency CV (pp)−7.0 ± 6.0−1.1 ± 6.0MD = − 5.9 pp (− 9.3, − 2.5)0.001SMD = − 0.93Rotation and system efficiencyRotation-shift latency (s)−0.10 ± 0.06−0.04 ± 0.06MD = − 0.06 s (− 0.092, − 0.028)0.001SMD = − 1.00Out-of-system rate (pp)−4.5 ± 3.8−1.3 ± 3.7MD = − 3.2 pp (− 5.2, − 1.2)0.002SMD = − 0.86≥ 20% OOS reduction, n/N (%)17/28 (60.7%)7/28 (25.0%)RD = 0.357 (0.118, 0.596)0.002RR = 2.43Defensive coverageCoverage index (0–1)+ 0.06 ± 0.05+ 0.02 ± 0.05MD = + 0.04 (+ 0.013, + 0.067)0.004SMD = 0.78Successful digs (pp)+ 7.0 ± 7.0+ 2.0 ± 7.0MD = + 5.0 pp (+ 1.2, + 8.8)0.010SMD = 0.71Coverage uniformity CV (pp)−6.0 ± 5.0−2.0 ± 5.0MD = − 4.0 pp (− 6.7, − 1.3)0.004SMD = − 0.78MD, mean difference in pre–post change between groups, calculated as intervention change minus control change. RD , risk difference. RR , risk ratio. SMD, standardized mean difference. OOS, out-of-system. pp, percentage points. Primary tactical outcomes. MD, mean difference in pre–post change between groups, calculated as intervention change minus control change. RD , risk difference. RR , risk ratio. SMD, standardized mean difference. OOS, out-of-system. pp, percentage points. Rotation-shift latency decreased by − 0.10 ± 0.06 s in the intervention group and − 0.04 ± 0.06 s in the control group, with MD = − 0.06 s (95% CI − 0.092, − 0.028; p = 0.001). SMD = − 1.00 indicates the strongest tactical timing advantage in this section. Out-of-system rate decreased by MD = − 3.2 pp (95% CI − 5.2, − 1.2; p = 0.002), indicating fewer disrupted offensive sequences after reception or defensive recovery. A ≥ 20% out-of-system reduction occurred in 60.7% of intervention athletes and 25.0% of control athletes, indicating a larger shift toward stable rally continuity in the intervention group. Defensive coverage index increased by + 0.06 ± 0.05 in the intervention group and + 0.02 ± 0.05 in the control group, with MD = + 0.04 (95% CI + 0.013, + 0.067; p = 0.004). SMD = 0.78 indicates a moderate-to-large intervention advantage in court-coverage change. Successful digs improved by MD = + 5.0 pp (95% CI + 1.2, + 8.8; p = 0.010), linking broader court occupation with more effective defensive contacts. Coverage variability decreased by MD = − 4.0 pp (95% CI − 6.7, − 1.3; p = 0.004), indicating more even defensive responsibility across court zones. Biomechanical execution In biomechanical execution, the RQ is answered by examining whether CNN-derived keypoint indicators improved jump performance, approach mechanics, arm-swing sequencing, and block-timing consistency. H2 is tested through between-group pre–post change analyses for jump height, take-off time, approach angle, step-length ratio, arm-swing lag, shoulder angular velocity, block-timing SD, sealing success, and predefined biomechanical responder thresholds. Jump height increased by + 5.7 ± 3.8 cm in the intervention group and + 2.7 ± 3.6 cm in the control group, with MD = + 3.0 cm (95% CI + 1.2, + 4.8; p = 0.002). SMD = 0.88 indicates a large intervention advantage in vertical displacement change, which corresponds to a higher attack reach and block contact position. Take-off time decreased by MD = − 9 ms (95% CI − 15, −3; p = 0.004), indicating faster jump initiation with higher jump output (Table S2). A ≥ 5 cm jump-height gain occurred in 64.3% of intervention athletes and 32.1% of control athletes, showing that visible height improvement occurred in twice the proportion of intervention athletes. Approach angle increased by + 2.2 ± 1.8° in the intervention group and + 0.8 ± 1.7° in the control group, with MD = + 1.4° (95% CI + 0.6, + 2.2; p = 0.001). SMD = 0.91 indicates a large intervention advantage in movement-path correction toward the 18–25° attack-lane target. Step-length ratio improved by MD = + 0.03 (95% CI + 0.01, + 0.05; p = 0.007), indicating better penultimate-to-final step organization during approach preparation. Trials inside the 18–25° band increased by MD = + 11 pp (95% CI + 4, +18; p = 0.003), showing more frequent execution inside the target approach window. Table 3Multiplicity-controlled primary outcomes.Primary outcomeMD (95% CI)pHolmSMDTransition timeMD = − 0.08 s (− 0.126, − 0.034)0.012−0.82Serve–receive accuracyMD = + 5.7 pp (+ 2.1, + 9.3)0.0120.87Rotation-shift latencyMD = − 0.06 s (− 0.092, − 0.028)0.009−1.00Out-of-system rateMD = − 3.2 pp (− 5.2, − 1.2)0.012−0.86Defensive coverage indexMD = + 0.04 (+ 0.013, + 0.067)0.0120.78Jump heightMD = + 3.0 cm (+ 1.2, + 4.8)0.0120.88Approach angleMD = + 1.4° (+ 0.6, + 2.2)0.0090.91Arm-swing lagMD = − 6 ms (− 10, − 2)0.012−0.81Block-timing SDMD = − 11 ms (− 17, − 5)0.009−0.95MD, intervention pre–post change minus control pre–post change; pHolm , Holm-adjusted p value; all nine primary outcomes remained significant under BH-FDR correction at q < 0.05. Table 3Multiplicity-controlled primary outcomes.Primary outcomeMD (95% CI)pHolmSMDTransition timeMD = − 0.08 s (− 0.126, − 0.034)0.012−0.82Serve–receive accuracyMD = + 5.7 pp (+ 2.1, + 9.3)0.0120.87Rotation-shift latencyMD = − 0.06 s (− 0.092, − 0.028)0.009−1.00Out-of-system rateMD = − 3.2 pp (− 5.2, − 1.2)0.012−0.86Defensive coverage indexMD = + 0.04 (+ 0.013, + 0.067)0.0120.78Jump heightMD = + 3.0 cm (+ 1.2, + 4.8)0.0120.88Approach angleMD = + 1.4° (+ 0.6, + 2.2)0.0090.91Arm-swing lagMD = − 6 ms (− 10, − 2)0.012−0.81Block-timing SDMD = − 11 ms (− 17, − 5)0.009−0.95MD, intervention pre–post change minus control pre–post change; pHolm , Holm-adjusted p value; all nine primary outcomes remained significant under BH-FDR correction at q < 0.05. Multiplicity-controlled primary outcomes. MD, intervention pre–post change minus control pre–post change; pHolm , Holm-adjusted p value; all nine primary outcomes remained significant under BH-FDR correction at q < 0.05. Block-timing SD decreased by − 17 ± 14 ms in the intervention group and − 6 ± 13 ms in the control group, with MD = − 11 ms (95% CI − 17, −5; p = 0.001). SMD = − 0.95 indicates the strongest biomechanical timing advantage, reflecting lower variability in block synchronization (Table 3). Sealing success increased by MD = + 7 pp (95% CI + 3, +11; p = 0.002), indicating that improved timing consistency produced more effective block closure. A ≥ 15% reduction in block-timing SD occurred in 57.1% of intervention athletes and 25.0% of control athletes, showing that timing consistency improved in more than twice as many intervention athletes. Fig. 2Biomechanical execution outcomes showing (A) jump-height change between the intervention and control groups, (B) approach-angle change plotted against step-length ratio change, (C) arm-swing lag change across backswing, acceleration, and contact phases, and (D) block-timing variability in the intervention and control groups. Biomechanical execution outcomes showing (A) jump-height change between the intervention and control groups, (B) approach-angle change plotted against step-length ratio change, (C) arm-swing lag change across backswing, acceleration, and contact phases, and (D) block-timing variability in the intervention and control groups. Arm-swing lag decreased by − 10 ± 9 ms in the intervention group and − 4 ± 8 ms in the control group, with MD = − 6 ms (95% CI − 10, −2; p = 0.004). SMD = − 0.81 indicates a large intervention advantage in proximal-to-distal sequencing (Fig. 2). Shoulder angular velocity increased by MD = + 55 °/s (95% CI + 20, +90; p = 0.003), indicating faster arm acceleration during the striking phase. Sequences within the ≤ 60 ms lag target increased by MD = + 13 pp (95% CI + 5, +21; p = 0.003), showing more frequent arm-swing timing inside the intended coordination window. Transfer & retention In transfer and retention, the RQ is answered by examining whether player-specific prescriptions improved within-session learning rate, scrimmage performance, and 1–2 week retention of tactical and biomechanical gains. H2 is tested through between-group pre–post change analyses for learning slopes, scrimmage outcomes, retention endpoints, and position-specific transition effects. Transition slope improved by − 0.022 ± 0.018 s/session in the intervention group and − 0.008 ± 0.017 s/session in the control group, with MD = − 0.014 s/session (95% CI − 0.022, − 0.006; p = 0.004). SMD = − 0.86 indicates a large intervention advantage in session-by-session tactical recovery speed (Table S3). Correct formation slope increased by + 1.35 ± 1.10 pp/session in the intervention group and + 0.45 ± 1.05 pp/session in the control group, with MD = + 0.90 pp/session (95% CI + 0.32, + 1.48; p = 0.004). SMD = 0.83 indicates a large intervention advantage in the rate of serve–receive organization improvement. Positive transition slopes occurred in 78.6% of intervention athletes and 53.6% of control athletes, while positive formation slopes occurred in 82.1% and 57.1%, respectively. Side-out rate increased by + 6.0 ± 5.0 pp in the intervention group and + 2.0 ± 5.0 pp in the control group, with MD = + 4.0 pp (95% CI + 1.1, + 6.9; p = 0.008). SMD = 0.72 indicates a moderate-to-large transfer advantage in serve–receive-to-attack continuity (Fig. 3). Point differential improved by + 1.5 ± 1.8 points/set in the intervention group and + 0.5 ± 1.9 points/set in the control group, with MD = + 1.0 point/set (95% CI + 0.3, + 1.7; p = 0.006). SMD = 0.75 indicates a moderate-to-large set-level advantage. Attack efficiency increased by + 4.5 ± 4.3 pp in the intervention group and + 1.2 ± 4.2 pp in the control group, with MD = + 3.3 pp (95% CI + 1.2, + 5.4; p = 0.004). SMD = 0.83 indicates a large offensive-efficiency transfer effect. Fig. 3Transfer and retention outcomes showing (A) intra-session transition-time learning curves across training sessions, (B) scrimmage performance changes in side-out percentage, point differential, and attack efficiency, (C) short-term retention of transition time from post-test to follow-up, and (D) position-specific transition-time gains across setters, middles, outsides, opposites, and liberos. Transfer and retention outcomes showing (A) intra-session transition-time learning curves across training sessions, (B) scrimmage performance changes in side-out percentage, point differential, and attack efficiency, (C) short-term retention of transition time from post-test to follow-up, and (D) position-specific transition-time gains across setters, middles, outsides, opposites, and liberos. Transition retention changed by + 0.02 ± 0.05 s in the intervention group and + 0.06 ± 0.06 s in the control group, with MD = − 0.04 s (95% CI − 0.07, − 0.01; p = 0.007). SMD = − 0.74 indicates a moderate-to-large retained advantage in tactical speed. Formation retention changed by − 1.5 ± 3.0 pp in the intervention group and − 4.0 ± 3.5 pp in the control group, with MD = + 2.5 pp (95% CI + 0.8, + 4.2; p = 0.005). Jump-height retention changed by − 0.8 ± 1.8 cm in the intervention group and − 2.6 ± 2.0 cm in the control group, with MD = + 1.8 cm (95% CI + 0.8, + 2.8; p = 0.001). SMD = 0.94 indicates a large retained advantage in vertical displacement. Approach-angle retention changed by − 0.3 ± 0.8° in the intervention group and − 1.0 ± 0.9° in the control group, with MD = + 0.7° (95% CI + 0.3, + 1.1; p = 0.001). SMD = 0.91 indicates a large retained advantage in attack-lane mechanics. Setters showed the largest position-specific transition effect, with MD = − 0.09 s (95% CI − 0.15, − 0.03; p = 0.014) and SMD = − 1.35, indicating a very large role-specific gain in offensive-tempo organization. Middles showed MD = − 0.09 s (95% CI − 0.15, − 0.03; p = 0.006) and SMD = − 1.08, indicating a large gain in block-to-attack transition speed. Outsides showed MD = − 0.08 s (95% CI − 0.14, − 0.02; p = 0.010) and SMD = − 1.02. Opposites showed MD = − 0.08 s (95% CI − 0.15, − 0.01; p = 0.031) and SMD = − 1.11. Liberos showed MD = − 0.06 s (95% CI − 0.12, − 0.01; p = 0.040) and SMD = − 0.98. Position-specific effects therefore showed large transition-speed gains across front-row and back-row roles, with the strongest practical gain in setters. Responder & heterogeneity analyses In responder and heterogeneity analyses, the RQ is answered by examining whether player-specific prescriptions produced threshold-level tactical improvement among athletes with different baseline performance profiles and prescription exposure levels. H2 is tested through prespecified responder analyses for transition time, serve–receive accuracy, rotation-shift latency, and coverage index; baseline-severity analyses for transition and serve–receive performance; dose–response analyses across prescription-exposure bands; and tactical–biomechanical concordance between transition time and jump height. Transition-time responders were 67.9% in the intervention group and 32.1% in the control group, with RD = 0.358 (95% CI 0.149, 0.567; p = 0.005), RR = 2.12, and NNT = 2.8. Approximately three athletes receiving the intervention produced one additional ≥ 10% transition-time responder compared with standard coaching (Table S4). Serve–receive responders were 60.7% and 32.1%, respectively, with RD = 0.286 (95% CI 0.071, 0.501; p = 0.012) and RR = 1.89. Rotation-shift responders were 53.6% and 25.0%, with RR = 2.14, while coverage-index responders were 50.0% and 21.4%, with RR = 2.34. Responder probabilities were therefore close to double across the main tactical domains. Baseline-severity analyses identified the largest transition-time gain in the slowest baseline quartile, with MD = − 0.11 s (95% CI − 0.17, − 0.05; p = 0.003). The lowest serve–receive quartile showed MD = + 8.0 pp (95% CI + 3.8, + 12.2; p = 0.009). The largest gains occurred in athletes with the weakest starting tactical profiles, indicating that the prescription rules were most responsive when baseline movement speed or formation stability was poorest. Prescription-exposure analyses showed larger tactical gains across higher exposure bands. Transition time improved by − 0.07 ± 0.06 s in the 0–3 prescription band, − 0.13 ± 0.07 s in the 4–7 band, and − 0.17 ± 0.08 s in the ≥ 8 band, with slope = − 0.050 s per exposure band (95% CI − 0.082, − 0.018; p = 0.003). Serve–receive accuracy increased by + 5 ± 6 pp, + 8 ± 6 pp, and + 11 ± 6 pp across the same exposure bands, with slope = + 3.000 pp per exposure band (95% CI + 1.020, + 4.980; p = 0.005). Tactical–biomechanical concordance was higher in the intervention group, where 46.4% improved both transition time and jump height compared with 17.9% of controls. κ = 0.46 and φ = 0.48 indicate moderate coupling between tactical speed and jump-performance gains in the intervention group. Transition time, approach angle, arm-swing lag, and block-timing SD had direct trigger thresholds, matched corrective drill families, and significant adjusted MDs in the linked outcomes (Table S5). The strongest timing pathway was block-timing SD, with MD = − 11 ms and q = 0.003, while the strongest tactical-speed pathway was transition time, with MD = − 0.08 s and q = 0.004 (Fig. 4). Fig. 4Responder and heterogeneity outcomes showing (A) responder-rate differences between the intervention and control groups, (B) baseline-quartile differences in transition-time change from Q1 to Q4, (C) dose–response patterns across prescription-exposure bands, and (D) concordance rates for combined tactical and biomechanical improvement. Responder and heterogeneity outcomes showing (A) responder-rate differences between the intervention and control groups, (B) baseline-quartile differences in transition-time change from Q1 to Q4, (C) dose–response patterns across prescription-exposure bands, and (D) concordance rates for combined tactical and biomechanical improvement. Usability and implementation findings showed how athletes and coaches interacted with the dashboard during live training. Athlete SUS was 78.6 ± 9.4 on the 0–100 scale, exceeding the 70-point acceptability threshold by 8.6 points. Athlete SUS was 78.6 ± 9.4 on the 0–100 scale, exceeding the 70-point acceptability threshold by 8.6 points. Feedback clarity was 4.3 ± 0.6, drill relevance was 4.1 ± 0.7, and feedback timing was 4.0 ± 0.6 on the 5-point scale, indicating that athletes rated the dashboard prompts as clear, task-relevant, and appropriately timed during dead-ball delivery. Training burden was 2.1 ± 0.8, below the scale midpoint, indicating low perceived disruption to practice. Coach dashboard readability was 4.4 ± 0.5 and drill-tile interpretability was 4.2 ± 0.6, indicating that coaches could read and interpret the prescription display during live sessions. NASA-TLX was 38.4 ± 8.7 on the 0–100 scale, placing coach workload below the midpoint of the scale. Prescription override rate was 6.8%, and mean tile-view duration was 7.4 ± 2.1 s, indicating brief dashboard interaction with most generated prescriptions delivered as displayed. Coding agreement reached κ = 0.84, exceeding the κ ≥ 0.80 threshold, and fidelity scores were closely matched across arms at 94.6 ± 3.1% in the intervention group and 93.8 ± 3.4% in the control group, with MD = + 0.8 pp. System validation, benchmarking, and generalization System validation answers the RQ by showing whether YOLOv5–DeepSORT–CNN produced accurate, real-time, player-specific data for training prescription. H1 is tested through athlete-independent detection, tracking, keypoint, reference biomechanical, latency, and OOD results. Computer vision validation supported H1 across detection, tracking, and keypoint outcomes (Table S6). Player mAP@0.50 reached 96.4%, with an MD of + 3.3 pp versus detector-only processing. Ball mAP@0.50 reached 91.8%, with an MD of + 4.2 pp, showing stronger small-object localization for the ball. IDF1 increased by + 11.8 pp and HOTA by + 12.2 pp, indicating more stable identity continuity and trajectory association during crossings, occlusions, and rally transitions. PCK@0.2 increased by + 8.1 pp, and joint-angle MAE decreased by 2.4° at the shoulder and 2.5° at the knee, giving the keypoint stream sufficient precision for movement-timing and kinematic extraction. Reference biomechanical validation showed acceptable agreement between video-derived and external reference measures (Table S7). Jump-height MAE was 2.1 cm, approach-angle MAE was 1.7°, arm-swing lag MAE was 9.6 ms, and block-timing SD MAE was 8.7 ms. ICC values from 0.79 to 0.88 indicate stable agreement for group-level kinematic analysis, with shoulder angular velocity showing the widest limits of agreement while remaining inside the prespecified MAE threshold. Latency benchmarks showed sustained real-time processing during full training sessions (Table S8). Full-pipeline latency was 69.9 ± 8.6 ms with median 68.4 ms (IQR 63.1, 74.6), and p95 latency remained 78.2 ms. Pose estimation accounted for 31.6 ± 6.5 ms, making it the largest runtime component. Throughput was 34.8 ± 3.1 fps, and dropped-frame rate was 0.42 ± 0.18%, keeping processing inside the ≥ 30 fps and ≤ 1% dropped-frame targets. OOD testing showed controlled performance loss across court, camera, and competitive-level changes (Table S9). mAP@0.50 decreased by − 3.7 pp, IDF1 by − 5.6 pp, HOTA by − 5.3 pp, and PCK@0.2 by − 3.6 pp from the in-distribution 4 K/60 fps condition. Each MD stayed within the prespecified OOD margin, so H1 remained supported outside the original court setup. ROI-cropped 1080p processing produced the strongest speed–accuracy balance, with p95 latency reduced to 46.8 ms and PCK@0.2 retained at 90.2%. Implementation diagnostics quantified the facility burden (Table S10). First-session setup required 52 ± 8 min, routine recalibration required 16 ± 4 min, and routine synchronization drift remained 4.6 ± 1.7 ms. Storage demand reached 410 ± 38 GB per 180 min session, making data management the largest practical burden. Operator interventions averaged 0.8 ± 0.6 per session, indicating that most sessions ran without repeated manual correction. The full 4 K/60 fps profile required an estimated startup cost of US$6,850–9,900, first-session setup of 52 ± 8 min, weekly recalibration of 16 ± 4 min, and storage demand of 410 ± 38 GB per 180 min session (Table S11). The ROI-cropped 1080p profile reduced the estimated startup cost to US$3,250–5,100 and lowered storage demand to 118 ± 22 GB per 180 min session, while keeping p95 latency at 46.8 ms. Coach-facing workload remained low during live practice, with tile-view duration of 7.4 ± 2.1 s and prescription override rate of 6.8%. The main implementation burden came from equipment purchase, video storage, calibration, and operator availability, while live coach interaction time remained brief. Discussion The study of 56 competitive volleyball athletes found comparable baseline characteristics between groups. The integrated YOLOv5–DeepSORT–CNN pipeline condition outperformed standard coaching on core tactical outcomes, delivering faster transitions, higher serve–receive accuracy with fewer reception errors, reduced within-set variability, quicker rotation shifts, fewer out-of-system plays, and stronger defensive coverage with more successful digs. Biomechanically, the pipeline yielded greater jump height, shorter take-off times, improved approach angles and step-length ratios, tighter arm-swing timing with higher shoulder angular velocity, and more consistent block timing with higher sealing success. Benefits transferred to scrimmage play, with better side-out rate, point differential, and attack efficiency, alongside steeper within-session learning and superior 1–2-week retention. Responder and dose–response analyses favored the pipeline, with stronger coupling between tactical and biomechanical gains, supporting durability. Automated visual measurement, threshold-based feedback, coach-delivered cues, and adaptive drill progression formed one intervention pathway. YOLOv5–DeepSORT–CNN supplied standardized estimates of player location, identity continuity, trajectory behavior, and keypoint movement during practice. Coaches used these estimates to deliver corrective drills when transition delay, spacing error, approach deviation, arm-swing lag, or block-timing variability appeared. Tactical gains indicate faster correction of spacing, rotation, transition, and coverage errors. Biomechanical gains indicate repeated correction of approach, jump, arm-swing, and block-timing patterns under live rally constraints. The integrated YOLOv5–DeepSORT–CNN condition outperformed standard coaching on core match-relevant metrics: transition time fell by an additional 0.08 s, serve–receive accuracy rose by an additional 5.7% points, rotation-shift latency decreased by an additional 0.06 s, out-of-system rate dropped by an additional 3.2% points, and defensive coverage index improved by an additional 0.04. These gains build on prior evidence that high-fidelity detection–tracking enables robust tactical extraction (MOTA 94.8; mAP 85.5%), but extend it from recognition to training effect by converting trajectories and keypoints into drill prescriptions1. Loureiro24, reported 15–20% faster GPAI improvements with flexible daily-adjusted planning; the rule-based adaptations provide an algorithmic analogue, yielding + 5.7% points in serve–receive and − 3.2% points in out-of-system plays over matched exposure. Mechanistically, the pipeline operationalizes a closed-loop, constraints-led approach: real-time state estimation (IDs, trajectories, pose) reduces decision latency and stabilizes team affordances; coupling tactical targets (spacing, timing) with biomechanical control variables (arm-swing sequencing, approach angle) tightens temporal coordination, lowering rotation and transition delays; and immediate drill selection functions as a feedback controller that diminishes error dynamics across rallies, producing durable improvements without increasing practice volume. The tactical analysis indicates a condition-level effect of the full YOLOv5–DeepSORT–CNN prescription workflow. Matched session volume, drill categories, rest structure, and coach–athlete ratio reduce the likelihood that the gains came from greater practice exposure. Component-level attribution remains limited because dashboard prompts, coach cueing, threshold-triggered drill selection, and progression rules were delivered as one active training condition. Automated measurement sharpened the timing and consistency of correction, allowing structured drills to target transition speed, formation stability, rotation timing, and coverage behavior during the same practice session. The pipeline condition produced biomechanical and positional gains beyond standard coaching: jump height rose + 5.7 cm versus + 2.7 cm (≈ 11% vs. ≈ 6%), take-off time shortened by an additional 9 ms, approach angle improved + 1.4° more, step-length ratio increased + 0.03 more, arm-swing lag decreased by an extra 6 ms with + 55 °/s higher shoulder angular velocity, and block-timing variability dropped − 11 ms with + 7 pp higher sealing success. Rada et al.,25 reported 7–25% jump gains from specialized plyometrics; the study ≈ 11% improvement emerged without a dedicated plyometric block, suggesting that error-triggered prescriptions coupled to pose–trajectory metrics can elicit strength–speed adaptations comparable to mid-range plyometric responses while also tightening timing control. Detection sensitivity to small/distant agents has been linked to superior multi-scale modeling (97.7% mAP with Shuffle Attention/DCNv3), aligning with the improvements in defensive coverage index (+ 0.04) and successful digs (+ 5 pp) that depend on accurate far-court tracking21. Mechanistically, adjusting approach angle and the penultimate-to-final step ratio increases horizontal braking impulse and tendon preload, amplifying stretch–shortening cycle efficiency; synchronizing arm-swing timing reduces proximal-to-distal phase lag, raising angular velocity; and stabilizing block-timing SD reflects enhanced predictive cueing to opponent kinematics, converting precise visual state estimates into repeatable motor solutions across rallies. The biomechanical gains align with motor-learning principles in which repeated task-specific correction reduces variability in movement execution. CNN keypoints made approach angle, step rhythm, arm-swing lag, and block-timing variability visible during practice. Neuromuscular change came from repeated volleyball actions and corrective drills, including approach-lane rebuilding, arm-sequencing work, and synchronized block timing. The pipeline sharpened measurement and drill selection, while the repeated corrected actions drove adaptation in jump production, arm acceleration, approach control, and block timing. At the execution level, the integrated YOLOv5–DeepSORT–CNN condition tightened timing and power transfer: arm-swing lag decreased an additional 6 ms (− 10 vs. − 4 ms), peak shoulder angular velocity rose an additional 55 °/s (+ 80 vs. + 25 °/s), block-timing variability fell an additional 11 ms (− 17 vs. − 6 ms), and block sealing improved by 7 pp; these effects persisted at 1–2 weeks with superior retention in approach angle (+ 0.7°) and jump height (+ 1.8 cm). High-accuracy biomechanics classifiers have shown that joint-angle patterns in arm swings and blocks are reliably detectable (98.58% accuracy), but they typically stop at recognition; the contribution translates recognition into corrective control by mapping pose-derived errors to drill prescriptions that compress the timing-error distribution and elevate proximal-to-distal velocity transfer17. Motivation and social context also shape learning dynamics motivation, coach–athlete relations, and cohesion jointly predict performance (R² = 0.71) and faster adaptation to corrections; the higher responder rate (67.9% vs. 32.1%) and positive dose–response slopes are consistent with a mechanism in which readiness modulates the feedback controller’s gain, increasing error sensitivity while lowering forgetting, thereby stabilizing motor solutions across rallies and into follow-up23. Scrimmage transfer shows that training gains reached a game-like practice context through side-out rate, attack efficiency, and point differential. Official matches add opponent scouting, crowd pressure, substitution strategy, referee flow, tactical deception, and score-state stress, which change the decision environment beyond controlled scrimmage play. Competition-level evaluation should track side-out percentage, attack efficiency, unforced errors, reception quality, block touches, transition attack success, and rally outcome under official match conditions. Match-based analysis would test whether the same tactical and biomechanical gains persist under opponent variability and competitive pressure. Theoretical and practical implications Theoretically, the study advances perception–action coupling by formalizing a closed-loop controller that links team-level trajectories with individual keypoint dynamics to drive adaptive prescriptions. Practically, the integrated YOLOv5–DeepSORT–CNN pipeline operationalized measurable targets (transition ≤ 1.20 s, approach 18–25°, lag ≤ 60 ms) at ≥ 30 fps and ≤ 80 ms latency, enabling within-session drill selection, progression/regression, and retention tracking without extra practice volume. Practical adoption fits collegiate programs, talent-development centers, and high-performance environments with stable camera placement, GPU access, operator support, and routine calibration capacity. The observed setup profile included four 4 K/60 fps cameras, GPU workstation, LED/NTP synchronization, first-session setup of 52 ± 8 min, weekly recalibration of 16 ± 4 min, and storage demand of 410 ± 38 GB per 180 min session, showing a moderate-to-high infrastructure requirement. The ROI-cropped 1080p profile reduced p95 latency to 46.8 ms while retaining usable tracking and keypoint accuracy, making it the most feasible lower-resource configuration for clubs with constrained hardware. Implementation value depends on video-review time saved, feedback consistency, drill-targeting precision, equipment cost, storage load, calibration time, and staff availability. Coach training demands appeared manageable because dashboard interpretation required short dead-ball interactions, with mean tile-view duration of 7.4 ± 2.1 s and prescription override rate of 6.8%. The main operating burden sits in setup, maintenance, data storage, synchronization checks, and operator availability. Programs with limited staffing or unstable court access can prioritize the 1080p ROI profile and a smaller metric set centered on transition time, approach angle, arm-swing lag, and block-timing SD. Limitations and future research Measurement, attribution, implementation, and competition transfer define the main boundaries of the study. Reference biomechanical validation was conducted in a subset, while full-session training relied on video-derived estimates; future work should extend synchronized motion-capture and force-plate validation across more sessions, positions, camera angles, and occlusion conditions. Component attribution remains limited because automated measurement, dashboard prompts, coach cueing, adaptive drills, and progression rules were delivered within one training condition. A three-arm design comparing standard coaching, adaptive drills without AI, and AI-supported adaptive drills would estimate the added value of automated vision beyond structured pedagogy. Official-match testing should include side-out rate, attack efficiency, unforced errors, reception quality, transition attack success, and rally outcome under competition constraints. Practical implementation should be evaluated through cost, setup time, storage burden, recalibration frequency, operator requirements, and coach-training demands across high-resource and lower-resource club environment. Conclusion The study indicates that the YOLOv5–DeepSORT–CNN workflow supported volleyball training by linking live trajectory tracking, keypoint estimation, threshold-based feedback, coach cueing, and adaptive drill progression within one practice system. Compared with standard coaching, the intervention condition showed faster transition execution, improved serve–receive organization, reduced out-of-system play, stronger defensive coverage, higher jump performance, faster take-off, improved approach mechanics, tighter arm-swing timing, and more consistent block timing. Transfer effects appeared in scrimmage side-out rate, point differential, and attack efficiency, while short-term retention favored the intervention condition for tactical speed, formation stability, jump height, and approach angle. Responder and dose–response patterns indicated stronger tactical gains among athletes receiving more targeted prescriptions, and concordance findings showed closer alignment between tactical and biomechanical improvement. Real-time video analytics therefore functioned as a measurement and prescription layer for individualized volleyball training in settings with sufficient camera, GPU, calibration, storage, and operator capacity. Policy implementation should prioritize standardized video-analytics benchmarks, coach certification in data-guided feedback, privacy-preserving athlete-data governance, minimum infrastructure specifications, and phased deployment models that allow high-performance centers and resource-limited clubs to adopt validated metrics at different levels of technical capacity. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1