icSPORTS 2026 Abstracts

Area 1 - AI, Data & Sport Systems
Full Papers Short Papers
Area 2 - Health, Sports Medicine & Rehabilitation Technologies
Full Papers Short Papers
Area 3 - Performance, Biomechanics & Coaching Support
Full Papers Short Papers

Area 1 - AI, Data & Sport Systems

Full Papers
Paper Nr: 19
Title:

A Predictive Framework for Evaluating Defensive Clears in Rocket League
Authors:

Undarmaa Baatarchuluun, Yumu Tabuchi and Keisuke Fujii
Abstract: Rocket League is a physics-based esport in which players control rocket-powered cars in continuous three-dimensional matches. Defensive clears are critical actions that influence short-term control and scoring risk, but are commonly evaluated using outcome-based measures such as goal concession, which confound situational difficulty with action quality. This paper proposes a predictive framework for evaluating defensive clears using short-horizon probabilistic modeling inspired by state-value approaches in sports analytics. Using 2,388 defensive clears extracted from 100 simulated 3v3 matches played by the Nexto bot, gradient-boosted models are trained to estimate probabilities of (i) winning the next ball, (ii) opponent immediate shot pressure, and (iii) opponent immediate goal. A dual-model formulation consisting of a pre-clear model and a transition-conditioned post-clear model is used to define ClearImpact, a descriptive measure of local game-state change following a clear. On a held-out test set of 761 clears, post-clear models achieve AUC values above 0.71 for win-next and opponent immediate shot prediction, and above 0.62 for goal prediction. ClearImpact aligns most strongly with opponent immediate shot outcomes. The proposed framework provides a structured, data-driven, and transition-aware approach for evaluating defensive actions in esports, while aiming to separate situational difficulty from local action effect.
Download
Paper Nr: 25
Title:

Quantifying Sports Form Proficiency Using Transfer Entropy-Based Body Coordination Network Analysis: Applied to Baseball Batting
Authors:

Takeshi Tanaka and Norio Gouda
Abstract: AI technology for image recognition has made motion analysis easier to implement, and it is expected to be widely used in areas such as exercise instruction and rehabilitation. However, interpreting the recognized body coordinate data requires specialized knowledge, and quantitative evaluation is a particular challenge. In this paper, we propose a method for quantifying motion characteristics by analyzing the network of whole-body coordination based on inter-region transfer entropy. Specifically, we propose a method for quantifying the number of edges and cliques connecting joints and the probability of occurrence of edges and cliques generation based on multiple coordination networks calculated continuously from a series of forms. Furthermore, we apply the proposed method to a prototype system and evaluate the batting forms of actual amateur baseball players. As a result, we quantify that the occurrence of edges and cliques based on the proposed indicators was concentrated on the coordination between specific joints due to the stabilization of forms with increasing skill, thereby confirming the validity of the proposed method.
Download
Paper Nr: 34
Title:

A Multi-Modal Action Spotting Transformer via Transfer Learning from Soccer to Futsal
Authors:

Yao Pan, Calvin Yeung and Keisuke Fujii
Abstract: Action spotting is the task of detecting and temporally localizing discrete events in continuous sports video streams. It underpins a range of downstream applications, including tactical analysis, referee assistance, and athlete training. Existing methods are predominantly designed for large-scale benchmarks and perform poorly when applied to data-scarce domains such as futsal. To address this, we extend the soccer action spotting model ASTRA to the futsal domain through transfer learning and a set of targeted design modifications, resulting in MMAST (Multi-Modal Action Spotting Transformer). The key modifications are: replacing ASTRA’s pre-extracted visual features with VideoMAEv2 for richer spatio-temporal representations; using a frozen pre-trained VGGish for audio rather than end-to-end training; adopting a 2-1-1 hierarchical encoder split that transitions to global attention earlier; and introducing a structured two-phase fine-tuning protocol in which the model is first pre-trained on SoccerNet and then progressively adapted to F-League futsal data. Experiments on 10 F-League matches showed that MMAST consistently outperformed the evaluated baselines under the transfer learning setting, achieving a Loose AmAP of 0.8641, 7.8 percentage points higher than the next best method. Ablation studies provided evidence for the contributions of the audio modality, the VideoMAEv2 visual backbone, and an encoder configuration with earlier global attention, which improved generalization to the target domain. These results suggest the potential of combining multi-modal pre-training with structured fine-tuning for action spotting when transferring to a closely related sport with limited annotated data.
Download
Paper Nr: 36
Title:

A Web-Based Tool for Localization Training in Orienteering from First-Person Video
Authors:

Marco L. Della Vedova
Abstract: In orienteering athletes are often required to localize themselves on a map while moving through terrain, but this skill is difficult to isolate and repeat in ordinary outdoor training. We present a web-based training application that converts first-person running video, GPS tracks, and orienteering maps into short localization tasks. In each task, the athlete observes a video segment, follows the movement on the map from a known start position, and submits an estimate of the final position after the clip ends. The application supports both multiplayer, gamified sessions and individual practice, and provides feedback based on distance error and response time. The associated dataset currently contains approximately 10.5 hours of first-person video from the Gothenburg area, with tasks curated from synchronized video, GPX, map, and metadata files. We describe the application concept, dataset workflow, and implementation, and report formative observations from 235 sessions with orienteers ranging from beginners to international competitive level.
Download
Paper Nr: 42
Title:

Late Fusion and Context-Free Pretraining for Cross-Context Expected Score Distribution Prediction in Digital Curling
Authors:

Rintaro Chiba, Yasumasa Tamura, Shimpei Aihara and Masahito Yamamoto
Abstract: In curling, the next-shot decision depends on the distribution of possible end scores, known as the Expected Score Distribution (ESD). The inner product of the ESD and a win probability vector (WV)—a tournamentspecific encoding of the game context—represents the expected win probability of a candidate shot. Since WVs differ across tournament categories and game formats, and retraining the predictor for every new WV is prohibitively expensive, the ESD predictor must generalize to unseen WVs. We compare three WV-injection architectures for a Transformer-based ESD predictor—early fusion (Concat), FiLM, and cross-attention— under random initialization and context-free pretraining, in which the encoder learns to predict the contextmarginalized ESD from a null WV. Evaluation uses a 5-fold leave-one-tournament-out protocol on four outof-distribution (OOD) conditions covering state, context, and format shifts. Late fusion outperforms early fusion on context-shift conditions; cross-attention with context-free pretraining is best everywhere, cutting mean KL by up to 32% relative to Concat (p<0.001, Wilcoxon). Early fusion keeps a small but significant edge only when the WV matches training (state-OOD)—a remaining gap toward a model robust on all OOD axes. We also identify an early-fusion pretrain failure: pretraining helps late fusion but hurts early fusion on format-level OOD—a conflict with the WV-in-encoder design.
Download
Paper Nr: 44
Title:

Environment-Aware Transformer for Muscle Activity Estimation in Sport Climbing
Authors:

Keita Kimura, Rintaro Chiba, Yasumasa Tamura, Shimpei Aihara, Satoshi Kato, Kaoru Kimachi, Daichi Yamashita, Mai Onaga and Masahito Yamamoto
Abstract: Estimating muscle activity, primarily measured via surface electromyography (sEMG), in sport climbing is essential for performance analysis and injury prevention, yet wearable sensors hinder natural movement. Visionbased methods that rely solely on skeletal kinematics fail to capture the varying muscular loads imposed by different hold configurations: identical joint angles can produce vastly different effort levels depending on the spatial arrangement of holds. We propose an environment-aware Transformer that estimates upper-body sEMG from monocular video by integrating 2D pose sequences with the pose-relative spatial layout of climbing holds via cross-attention. We compare the proposed Hybrid model against kinematics-only baselines—both monocular 2D pose and a six-camera 3D markerless motion capture system—across N = 30 random seeds with paired t-tests and Cohen’s d. The Hybrid model attains the best amplitude accuracy (R 2 = 0.238, r = 0.706, RMSE = 9.11%Peak on held-out test trials), a statistically significant improvement over the kinematics-only pose baseline (R2: p < 0.001, d = 0.73; RMSE: p < 0.001, d = −1.28). At this level the predicted envelopes already track the coarse temporal and relative-magnitude structure of muscle load, suggesting potential utility for effort- and load-trend analysis, even though the generalisability of this accuracy beyond the two study participants remains largely unverified. Individual hold-integration components improve amplitude without significantly improving the Pearson correlation coefficient, revealing that hold context primarily corrects activation magnitude rather than temporal waveform shape. These results suggest that, for waveform fidelity and amplitude error, augmenting a monocular system with environmental context can match or exceed the benefit of multi-camera 3D depth.
Download
Paper Nr: 46
Title:

Markerless Rotation Estimation of Curling Stones from Monocular Video Using Two-Stage Handle Detection
Authors:

Takumu Nakamura, Shimpei Aihara and Takeshi Ito
Abstract: In curling, stone rotation affects the trajectory of the stone and consequently influences shot accuracy and game outcomes. As the strategic use of rotation has become increasingly important, a markerless method for measuring stone rotation is required. This study proposes a monocular vision-based approach for estimating stone rotation without markers. The proposed method employs a two-stage detection framework to accurately detect the stone handle, transforms image coordinates into world coordinates through camera calibration, and estimates stone rotation by tracking changes in handle orientation over time. The method was evaluated on 12 trials covering two stone colors, two rotation directions, and three rotational-speed conditions using motion-capture-based ground truth. Experimental results demonstrated that the proposed two-stage method substantially improved handle-angle and rotation-count estimation compared with single-stage detection. The proposed method achieved a handle-angle MAE of 3.813º and rotation-count MAE of 0.017 revolutions across 12 trials.
Download
Paper Nr: 66
Title:

Data-Driven Pose-Aware Team-Space Evaluation in Soccer with 3D Body Orientation
Authors:

Bin Liu and Keisuke Fujii
Abstract: Evaluating space is one of the fundamental aspects of contemporary soccer analytics. However, most reproducible models predominantly depend on player locations while neglecting body orientation, a tactical cue that is challenging to acquire on a large scale. This gap motivates a practical test of whether 3D body orientation can support data-driven pose-aware team-space evaluation under realistic linked-data quality. We propose a pose-aware evaluation framework that links PFF FC event-position data with WorldPose 3D body orientation and evaluates it through pass-success probability estimation against a location-only SoccerMap baseline. The framework quantifies three data-quality ceilings: no pose token for 70.4% of events, a 7.61 m median pose-to-pitch error, and a 75% identity reassignment rate. A controlled event-level probe further shows that facing the goal is positively associated with pass success. However, adding orientation does not improve system-level ranking over the SoccerMap baseline, and the apparent Log Loss gain is reproduced by shuffled pose, indicating calibration shrinkage rather than orientation content. These results indicate that body orientation contains interpretable, event-level tactical information, but the current pose representation gives no stable system-level gain beyond the strong location baseline.
Download
Paper Nr: 68
Title:

Can Vision-Language Models Judge Olympic Diving? From Reasoning to Scores in Zero-Shot Action Quality Assessment
Authors:

Henry O. Velesaca, David Freire-Obregón, Luigi Miranda and Abel Reyes-Angulo
Abstract: Automated action quality assessment (AQA) in Olympic sports remains a challenging task due to the complexity of human motion and the subjectivity inherent in expert judging. This work evaluates the capability of open-source Vision-Language Models (VLMs) to perform zero-shot action quality assessment on Olympic diving videos using the AQA-7 benchmark dataset. In this regard, a regression-based framework is proposed to leverage both the semantic reasoning and phase-level sub-scores generated by the VLMs, combining TF-IDF vectorization, dimensionality reduction, and ensemble learning to predict final competition scores. Experimental results show that standalone VLMs achieve moderate Spearman correlations below 0.32, while the proposed ensemble regression framework substantially improves performance in the reported evaluation, reaching a Spearman correlation of 0.67 with a four-model configuration. Textual reasoning features consistently outperformed raw numerical sub-scores, highlighting the richness of VLM-generated explanations for action quality analysis. These findings suggest that VLMs hold strong potential as assistive tools for explainable and semi-automated sports performance evaluation. The code is publicly available on GitHub https://github.com/hvelesaca/olympic diving judge vlm.
Download
Short Papers
Paper Nr: 24
Title:

Squat Signatures: Athlete Identification from a Single Wrist-Worn Sensor
Authors:

Bastian Dänekas and Rainer Malaka
Abstract: Individual movement patterns in human motion are well established, yet their exploitation in resistance training remains largely unexplored. We investigate whether a single wrist-worn smartwatch is sufficient to reliably identify athletes during squats based solely on their personal movement signatures. Using the BigThree100 dataseta which contains squat inertial measurement unit (IMU) data from 98 athletes (four sets each, 10,069 squat repetitions) we extract 63 features from acceleration and rotation magnitudes that capture movement style, coordination, smoothness, and spectral characteristics while reducing dependence on sensor orientation and range-of-motion cues. A Histogram Gradient Boosting classifier is trained in a leave-one-set-out setup. With just one repetition observed, identification accuracy already exceeds chance by a factor of approximately 31 (30.98% vs. 1.02%). With 15 repetitions the model reaches 85.65%, and with 30 consecutive repetitions 92.26%. A rarefaction analysis shows that accuracy degrades only gradually as population size grows to 96 athletes, while the improvement factor over random guessing continues to rise beyond that point. Feature importance analysis reveals normalized signal variance as the dominant predictor, and mismatches between athletes are more consistent with set-level variability than with stable pairwise similarity between athletes. These results demonstrate that stable person-specific movement signatures are present in smartwatch squat data and motivate future work on individuality-aware models for exercise execution analysis.
Download
Paper Nr: 27
Title:

Predicting Curling Stone Breakpoints from Early Trajectory Observations
Authors:

Guanyu Chen, Shimpei Aihara and Yoshinari Takegawa
Abstract: Curling stone trajectories provide cues for understanding shot behavior, but the breakpoint at which a stone begins sustained deviation from the delivery line is difficult to define and forecast automatically. We define the breakpoint as an operational trajectory event using delivery-line coordinates and noise-adaptive residual onset, and predict its longitudinal location from the first five seconds of motion. Using 525 quality-controlled practice trajectories, we compare regression and sequence models under a session-wise split. The compact GRU yielded the lowest point estimate among the evaluated models for timing-equivalent MAE (0.57 s) on the held-out test split. These results indicate that early two-dimensional trajectories contain information about the later operational breakpoint, although expert label validation and evaluation across additional sessions and rinks remain necessary.
Download
Paper Nr: 33
Title:

Comparative Analysis of Bottom-Up 2D Pose Estimation Methods for 3D Human Pose Estimation in Real-World Rugby Impact Scenarios
Authors:

Shi Yuchen, Ozeki Nobutake, Inaji Motoki, Yagishita Kazuyoshi and Miyazaki Yusuke
Abstract: In video-based sports biomechanics analysis, multi-stage 3D human pose estimation frameworks are widely used to quantify athletes' kinematics from monocular video. Within such frameworks, the 2D pose estimator constitutes a critical bottleneck determining final 3D accuracy. However, the impact of different 2D estimation methods on 3D performance in complex real-world sports scenarios has not been systematically evaluated. Using head impact events in rugby matches as the test scenario, this study integrates three generalized bottom-up 2D pose estimation methods — Associative Embedding (AE), Disentangled Keypoint Regression (DEKR), and Contextual Instance Decoupling (CID) — into a unified multi-stage 3D pose estimation framework, and conducts a comparative analysis using three metrics: MPJPE, P-MPJPE, and MPJVE. Experimental results show that AE performs significantly worse than both DEKR and CID across all metrics (p < 0.001), with its reliance on post-processing grouping exhibiting notable fragility in scenarios involving dense athlete contact and severe occlusion. DEKR and CID show no statistically significant difference in absolute localization accuracy (p = 0.796). However, DEKR significantly outperforms CID in pose structural accuracy and temporal smoothness (p < 0.05). This study provides empirical evidence to guide 2D estimator selection when constructing 3D pose estimation systems for complex sports scenarios.
Download
Paper Nr: 41
Title:

Motion-Relation Modeling for Skeleton-Based Tactical Action Localization in 3-on-3 Basketball
Authors:

Zihan Ye and Yoshinari Kameda
Abstract: Basketball tactical action localization requires temporal boundary estimation and reasoning over coordinated player interactions. We present a motion-relation framework that represents players as actor tokens and ordered pairs as directed relation tokens. A GATv2-style encoder aggregates dynamic interactions, while a gated ActionFormer-style head localizes tactical categories and boundaries. We also define Primary-4, a core-action benchmark, and Coarse-4, which maps all 19 annotations to four tactically motivated functional anchors. Across three seeds, the canonical Primary-4 model achieves 58.37 ± 1.98% mAP@0.5 and 26.25 ± 0.89% mAP@[0.50:0.05:0.95]. Matched ablations support temporal gating, whereas removing the role proxy does not reduce accuracy. These results establish a compact protocol for skeleton-based tactical localization while exposing limitations in strict boundary alignment, split verification, and role interpretation.
Download
Paper Nr: 43
Title:

A Learning-Based Stone Identification Method toward Building a Tactical Analysis Framework for Curling
Authors:

Kei Suzumura, Yasumasa Tamura, Shimpei Aihara and Masahito Yamamoto
Abstract: In curling tactical analysis, accurate stone tracking across shots is one of the essential pieces of information. Existing rule-based methods formulate the stone identification problem as a linear assignment problem using Euclidean distance as the cost function with physical constraints specific to curling. While effective for simple shots, these methods struggle with take-out shots such as Promotion Take-out and Hit and Roll, where multiple stones move simultaneously and the shooter often stops near the position of a displaced stone. This study proposes a learning-based approach that replaces the distance-based cost function with a multi-layer perceptron (MLP) trained using the Sinkhorn algorithm as a differentiable assignment operator. Experiments on 93 annotated ends demonstrate improvements of + 1.08 % in take-out shot transition accuracy and + 7.10 % in complete tracking rate over the rule-based baseline. A comparison with a linear cost function confirms the improvement stems from the model’s nonlinearity. This improvement, while statistically significant on average, varies in magnitude across random initializations, reflecting the limited size of the current dataset.
Download
Paper Nr: 45
Title:

Does the "Good Length, Outside Off" Wicket Hotspot Survive Match Context? A Calibrated Ball-Tracking Model for Twenty20 Cricket, with a Recency Null and a Validated Where-to-Bowl Tool
Authors:

Yuvaan Pandey
Abstract: Conventional cricket wisdom holds that the most dangerous place to bowl in Twenty20 is a good length, just outside off stump, and aggregated ball-tracking data appear to agree. But deliveries to that region are not bowled at random: they coincide with particular bowlers, phases and match situations. We ask whether the good-length/outside-off wicket hotspot survives once match context is controlled, using official Hawk-Eye ball-tracking data and a probabilistic next-ball wicket model: a logistic generalized additive model with a tensor smooth over pitch location plus leakage-free situational state, embedded in a nested ladder (locationonly, context-only, full) under match-grouped cross-validation. On 135,174 men’s IPL deliveries, location alone reaches Average Precision 0.056 against a 0.047 base rate, context alone 0.076, and the full model only 0.081; the hotspot’s held-out wicket rate is just 1.13× the base rate. The hotspot is real but largely a proxy for context, replicating on 213,078 ODI deliveries. We further show that recent ball-by-ball history is a predictive null, whereas the calibrated surface yields an honestly-validated, support-constrained where-tobowl tool whose recommended zones carry materially higher observed wicket rates.
Download
Paper Nr: 58
Title:

From AI-Supported Social Media Monitoring to Athlete Online Safety Governance: A Comparative Policy Review for Sport Information Systems
Authors:

Nagayuki Saito
Abstract: Online abuse against athletes has become a sport information systems problem as well as a welfare and governance problem. This position paper examines how AI-supported social media monitoring can be connected to organizational workflows, platform reporting, legal escalation, mental health support, and athlete education. Drawing on a rapid comparative policy review, the paper compares AI-supported initiatives by FIFA, the IOC, and World Athletics with policy models in Australia, the United Kingdom, and the European Union, and examines current Japanese arrangements as the setting to which these lessons are applied. The analysis shows that automated detection is useful but insufficient unless embedded in a socio-technical safeguarding system. The paper proposes an Athlete Online Safety System for Japan, consisting of cross-sport risk governance, AI-supported monitoring, human review and triage, trusted reporting to platforms, evidence preservation and legal escalation, and athlete support covering exposure control, psychological support, and scenario-based digital resilience education. The contribution is to reposition sport data science beyond performance enhancement and toward the protection of athletes in digitally mediated sport environments.
Download
Paper Nr: 63
Title:

Search-Based Policy-Value Learning with Game-State Augmentation for Digital Curling AI
Authors:

Tomomichi Kanaeda, Shimpei Aihara and Takeshi Ito
Abstract: Digital curling AI must handle computationally expensive stochastic simulations and strategic decisions that depend on score difference, remaining ends, and hammer possession. This paper proposes a search-based policy-value learning pipeline for the third-generation Digital Curling System (DC3). A Transformer-based network predicts shot and 17-class end-score distributions. SHOT generates target policy and score distributions, and reverse-curriculum learning proceeds backward from the 16th to the 10th shot of the 10th end. Game-state augmentation reassigns the end indices of states extracted from different ends, increasing the number of training states from approximately 15,000 to 150,000 without requiring additional game logs. During play, the learned distributions are combined with an empirical win-probability table and PUCT search. The system was evaluated on 1,000 states per shot stage using paired matches (2,000 games per condition). From 10th-shot states, it achieved mean score rates of 57.08% against KuraCurling_SHOT and 54.73% against a supervised-learning model using the same PUCT search. Ablation results showed advantages of 2.45 and 11.03 percentage points for value-network-based state evaluation and game-state augmentation, respectively, while no statistically significant difference was found between PUCT and SHOT. Among the tested components, game-state augmentation produced the largest observed improvement under the evaluated 10th-shot condition.
Download
Paper Nr: 64
Title:

Towards Context-Aware Expected Assist Estimation in Football: A Clustering-Based Label Smoothing Framework with Preliminary Two-Stage Validation
Authors:

Dhruv Manoj and Badri Prasad
Abstract: Expected Assists (xA) quantify a passer’s contribution to goal-scoring opportunities and have become a central metric in modern football analytics. The prevailing approach assigns xA directly from the Expected Goals (xG) value of the resulting shot, treating all shot-assist passes as equivalent provided the shot quality is similar. This conflates two analytically distinct quantities: the probability that a pass generates a shot at all, and the quality of the shot that follows when it does. We propose a clustering-based label smoothing framework that groups shot-assist passes by spatial and contextual similarity using DBSCAN, and assigns each pass a smoothed pseudo-xA label as a weighted blend of cluster-level mean xG (70%) and pass-length-weighted individual xG (30%). A fully connected neural network trained on these labels achieves an RMSE of 0.0577 on held-out data, a 77.1% reduction relative to a raw-xG baseline (RMSE = 0.2515); because the pseudo-xA label is itself partly constructed from cluster means, this reduction chiefly reflects the internal consistency of the smoothing procedure rather than externally validated predictive gain, and a cluster-mean-only baseline attains a comparable RMSE of 0.0592. To assess the empirical tractability of the natural two-stage extension this work motivates, we conduct preliminary experiments on the full StatsBomb open-data corpus of 3,836,550 passes across 80 competition-seasons. A gradient-boosted classifier for P(shot | pass) achieves AUC = 0.950 with strong calibration (Brier = 0.015) across the probability range containing 99.99% of test observations, and a separate regressor for E[xG | shot,pass context] achieves r = 0.745 against held-out shot outcomes. These preliminary results indicate that a full two-stage model is empirically tractable at scale, though joint calibration and independent cross-validation against ground truth remain open for future work. We integrate the label-smoothing framework into an interpretable player rating application and discuss its implications for crediting creative contributions presently invisible to conventional xA
Download
Paper Nr: 65
Title:

Tactical Evaluation of Tennis Strokes via Opponent Next-Swing Region Prediction from Broadcast Videos
Authors:

Koki Nakaseko, Kenjiro Ide, Ning Ding and Keisuke Fujii
Abstract: This paper tactically evaluates tennis strokes using broadcast-video tracking data and pose estimates. This is not biomechanical stroke-quality assessment: the action is the court region where the opponent next swings after the current shot, not the ball landing point. A recurrent model estimates action values for 15 regions from fixed logged rallies and is trained with a state–action–reward–state–action (SARSA)-style temporal-difference (TD) objective and supervised action imitation. We compare tracking-only and pose-based inputs; pose representations comprise raw coordinates, a multilayer perceptron (MLP), and a graph neural network (GNN) using a human-skeleton graph. Experiments use four US Open finals and evaluate held-out rallies with losses, classification and distribution metrics, and court heatmaps. For four players sharing the common setting, MLP encoding achieves the lowest mean total loss (0.443) and action loss (0.888), whereas GNN encoding achieves the lowest mean TD loss (0.155). These case-study results indicate that pose can help in some settings, but the preferred representation depends on the player and on whether the objective is next-swing-region prediction or logged-trajectory value consistency.
Download
Paper Nr: 87
Title:

From Match Momentum to Player Impact: An Interpretable Event-Based Framework for Football Player Ranking
Authors:

Michał Koczkodaj and Tomasz Piłka
Abstract: Evaluating football players from event data requires combining multiple dimensions of on-ball performance while preserving the transparency of the final ranking. This study proposes an interpretable event-based framework for ranking player–match performances using all 64 matches of the 2018 FIFA World Cup. Recorded actions are aggregated into four performance components: attacking contribution, offensive build-up, player involvement, and defensive contribution. These components are combined with player-level momentum and match-result context in a decomposable Player Impact Score. Event-derived variables are adjusted for estimated playing time where appropriate and normalised within each match. The analysis covers 1,787 player– match records, including 1,521 observations with at least 30 estimated minutes. The resulting rankings are examined through match-level momentum visualisation, score decomposition, a sensitivity analysis of the component weights, and an unsupervised PCA-based reference ranking, which agrees closely with the proposed score (median within-match Spearman correlation 0.97). The framework provides a reproducible and transparent description of recorded player contributions and should be interpreted as a descriptive ranking framework rather than a validated estimate of overall player quality.
Download
Paper Nr: 88
Title:

Automated Video Collection for Baseball Hitting Mechanics Analysis
Authors:

Ami Morizumi and Ryusuke Miyamoto
Abstract: Baseball hitting mechanics analysis requires side-view videos in which the hitter’s full body is visible throughout the swing. However, baseball broadcast videos contain irrelevant scenes, such as pitching, fielding, camera-angle changes, and close-up shots, making automatic collection difficult. This paper proposes an automatic method for extracting side-view swing scenes from baseball broadcast videos. The proposed method first uses CLIP to coarsely extract candidate swing segments. The candidates are then filtered and verified using HSV-based camera-transition detection, CLIP-based action filtering, viewpoint verification based on CLIP image features. Experiments were conducted using ten publicly available baseball videos totaling approximately 3.9 hours. The proposed method extracted 172 video segments, of which 137 were verified as valid swing scenes, achieving a precision of 79.7%. These results demonstrate that the proposed method can automate the collection of side-view swing videos.
Download
Paper Nr: 56
Title:

Comparative Analysis of Deep Learning Architectures for Exercise Classification Using Wrist-Worn IMU Data
Authors:

Harijs Tekutis, Jevgenijs Telicko and Zane Ruza
Abstract: Accurate monitoring of physical exercise is important for assessing training load, supporting adherence, and enabling data-driven feedback in fitness, rehabilitation, and health applications. This study investigates the effectiveness of classical machine learning and deep learning methods for exercise classification using wrist-worn inertial measurement unit (IMU) data. Using the RecGym dataset, which contains 11 gym exercises and a Null class, we compared classical ML methods with several hybrid CNN–RNN architectures based on BiLSTM and BiGRU layers, including residual and multi-head variants. Experimental results on the fixed subject-independent RecGym split showed higher mean macro F1 scores for the evaluated deep learning models than for the classical baselines. Among the tested configurations, CNN ResBiGRU achieved the highest mean test macro F1 score.
Download
Paper Nr: 60
Title:

Agents in the Soccer Booth: Can LLMs Follow Live Soccer Narratives?
Authors:

Vitor Rolla, Thamiris Neves, Luiz Carlos Marques, Gabriel Freire, Aline Cardoso, Gabriel Rodrigues and Lester Faria
Abstract: This demonstration paper evaluates whether modern large language models can reliably follow soccer narratives, reconstructing the final score and information surrounding the goals that occur during the match. The increasing adoption of generative artificial intelligence for task automation has intensified the need for reliable structured information extraction. Although newer and larger models are often assumed to provide superior structured outputs, empirical evidence under realistic conditions remains limited. In this work, we present a controlled benchmarking study of structured goal-event extraction from live soccer match commentary. We evaluate eight production-grade models from two major providers across three prompt formulations that represent different extraction philosophies. Performance is measured using Exact Match Rate under Relaxed and Strict reconstruction protocols, complemented by goal-level coverage analysis. Results show that prompt design significantly influences robustness. Under the Relaxed Exact Match protocol, the best-performing model exhibits a 38.8 percentage-point variation across prompt formulations. Notably, larger or more recent models, such as gemini-3.1-pro-preview, do not uniformly achieve superior Strict reconstruction performance. Our findings highlight the gap between entity recognition and full event grounding, and demonstrate that structured reliability depends not only on model scale but also on prompt strategy and evaluation rigor. The demo video is available at https://youtu.be/RlaRQrty4bs. The app is publicly available at https://sn-narratives.streamlit.app. The dataset and the codea are available at https://github.com/vitor-rolla/sn-agent.
Download

Area 2 - Health, Sports Medicine & Rehabilitation Technologies

Full Papers
Paper Nr: 23
Title:

Prospective Evaluation of Unilateral Shoulder Range of Motion and Functional Performance as Predictors of Noncontact Upper Extremity Injury in Overhead Athletes
Authors:

Bayram Bilgi, Hidayet Beyhan and İrem Düzgün
Abstract: Background: Preseason identification of modifiable risk factors for noncontact upper extremity injury is a clinical priority in overhead sports. Objective: To determine whether preseason unilateral shoulder range of motion (ROM) and functional performance measures are independently associated with noncontact upper extremity injury in overhead athletes. Methods: In this prospective cohort study, 136 overhead athletes (handball n=52; volleyball n=84; 79 females, 57 males) underwent preseason assessment and were followed for two competitive seasons. Glenohumeral internal rotation (IR), external rotation (ER), and total rotational ROM (TROM) were measured bilaterally. Functional performance was assessed with the Upper Quarter Y-Balance Test (UQYBT), Upper Limb Rotation Test (ULRT), Unilateral Seated Medicine Ball Throw Test (USMBT). LASSO logistic regression with bootstrap internal validation was used to identify independent predictors. Results: Twenty-two noncontact upper extremity injuries were recorded. Shoulder ER (OR=0.95; 95% CI: 0.91–0.99; p=0.015) and IR (OR=0.95; 95% CI: 0.91–0.99; p=0.019) were independently associated with injury risk. Adjusted for ROM, female sex was the strongest predictor (OR=4.64; p=0.014). No functional performance test was independently associated with injury. Optimism-corrected AUC was 0.61 (95% CI: 0.47–0.73). Conclusions: Absolute unilateral glenohumeral rotational ROM, but not functional performance, was independently associated with noncontact upper extremity injury risk in overhead athletes.
Download
Short Papers
Paper Nr: 40
Title:

Acute Flexibility and Balance Responses to Whole-Body Vibration in a Sham-Controlled, Fixed-Order Design: Null Findings for Vibration-Specific Effects in Sedentary Young Adults
Authors:

Abdalla Mohamed Abdalla and In-Ju Kim
Abstract: Sedentary office work rarely leaves room for structured exercise. We asked whether Whole-Body Vibration (WBV), delivered in under five minutes, produces acute gains in lower-body flexibility and standing balance in sedentary young adults. Forty inactive adults (N = 40; 28.07 ± 4.86 years; 75% male; IPAQ-inactive) completed three conditions in a fixed within-subject order: Sham Standing, WBV Standing, then WBV Semi-Squat; all at 30 Hz with three 60-second bouts. Sit-and-reach, single-leg stance, and forearm plank hold were measured at baseline and after each condition. Sit-and-reach improved across the sequence (F(2.45, 95.67) = 19.80, p < .001, η²p = .336); the absolute gain from baseline to semi-squat was 4.2 cm (d = 1.21), and the Sham vs. Semi-Squat difference, confounded by posture and order, was 1.9 cm (d = 0.48). The posture-constant contrast (Sham vs. WBV Standing) was non-significant (p = .136). Single-leg stance was unreliable (Spearman's ρ = 0.43) and is reported descriptively; plank findings (F(2.32, 90.60) = 1.78, p = .169) were inconsistent across analyses and were not interpreted as evidence of improved core endurance . Because the fixed order confounds vibration, posture, and session sequence, no vibration-specific benefit can be claimed. Within-subject sham controls should become standard in acute-WBV research, and counterbalanced replication is required before any workplace-deployment claim.
Download
Paper Nr: 52
Title:

Associations between Exercise Motives and Weekly Running Distance among Recreational Female Runners: A Cross-Sectional Study

Authors:

Danijela Kuna
Abstract: This cross-sectional study examined associations of exercise motives, age, and running experience with selfreported weekly running distance among recreational female runners. The sample comprised 304 Croatian women aged 18 −70 years (M = 36.71, SD = 10.36), who reported 24.34 ± 16.25 km per week and completed the Exercise Motivations Inventory-2 (EMI-2). Pearson correlations and hierarchical multiple regression were performed, with age and running experience entered first and the 14 EMI-2 dimensions added in the second step. Because heteroscedasticity was detected, HC3 robust standard errors were used. The model explained 27.8% of the variance in weekly running distance, with the EMI-2 dimensions accounting for an additional 17.1% (p < .001). Older age, greater running experience, enjoyment (B = 5.26, p = .004), challenge (B = 4.06, p = .036), and affiliation (B = 2.75, p < .001) were positively associated with weekly running distance. Positive health showed a negative conditional association (B = −5.90, p = .014), despite a nonsignificant zero-order correlation, suggesting a possible suppression effect. Weekly running distance was therefore associated with demographic, experiential, intrinsic, competence-related, and social factors. Longitudinal research using objective running measures may clarify how these associations develop over time.
Download
Paper Nr: 80
Title:

Wrist Is Not Chest: Robust Heart-Rate Estimation from Optical Sensors During Running and the Cost of Subject-Dependent Evaluation
Authors:

Ashwath Ram
Abstract: Optical heart-rate sensing at the wrist underlies consumer running watches, yet the arm motion intrinsic to running corrupts the photoplethysmography (PPG) signal it depends on. This is an evaluation and robustness study rather than an algorithmic contribution. We study heart-rate estimation on the public IEEE Signal Processing Cup 2015 dataset: twelve subjects run on a treadmill at varying speeds while a wrist unit records two PPG channels and tri-axial acceleration against a chest-ECG reference. Two findings emerge. First, the practical failure of simple methods is not high average error but occasional catastrophic divergence: an accelerometer-aware spectral tracker attains a median error of 3.5 beats per minute (BPM) yet a mean of 13.4 BPM, because on three of twelve subjects it locks onto the stride cadence. A compact gradient-boosted regressor on label-free spectral and motion features, evaluated leave-one-subject-out (LOSO), trades a higher median (9.0 BPM) for a large reduction in these failures, giving a 9.1 BPM mean and near-zero bias; robustness, not peak accuracy, is what a wearable should optimise. Second, the evaluation protocol is itself a confound: a random-window split that allows a subject’s windows into both folds scores the same model at 5.1 BPM, understating the LOSO error by 44% on this dataset. We quantify that optimism, and report the full per-subject error distribution so that the stability of these comparisons across only twelve subjects can be judged directly.

Area 3 - Performance, Biomechanics & Coaching Support

Full Papers
Paper Nr: 38
Title:

An Analysis Support System for Dynamic and Spatial Differences between Motions via Interactive Planar Visualization
Authors:

Daisuke Soejima, Kunio Ymamoto and Masaki Oshita
Abstract: In sports and skill training environments, there is an increasing demand to compare and analyze two different motions, such as those of an expert and a learner, to provide feedback for skill improvement. However, conventional visualization methods, such as side-by-side video presentation, subjective observation, or simple 3D trajectory displays, make it difficult to intuitively understand detailed differences in motion quality and the issue causes producing those differences. In this study, we propose a visualization analysis support system that transforms 3D motion data into a 3D density space (voxel space) with spatial extension, allowing users to interactively slice and observe it on a cross-section. By calculating and maintaining features such as spatial occupancy, motion velocity, and rotation amount, multi-faceted comparisons are enabled. Through experiments, we demonstrated that this system allows users to intuitively understand the interlocking of multiple body parts and differences in three-dimensional spatial relationships, whereas conventional methods were limited to local observations.
Download
Paper Nr: 69
Title:

Stroke Analysis in Wheelchair Tennis Based on Wheelchair Recognition and Skeleton Estimation
Authors:

Heng Zhang and Yoshinari Kameda
Abstract: Objective performance analysis in para-sports is essential for evidence-based training, yet sensor-based approaches in wheelchair tennis face practical limitations and cannot fully capture the complex interactions among the wheelchair, player, and ball. This study proposes a comprehensive, non-contact stroke analysis framework based on single-view match video. First, a novel three-dimensional wheelchair localization method is introduced to estimate precise court coordinates. Wheel regions are detected using RF-DETR, and perspective correction is applied to project the wheelchair position into the court plane, overcoming inaccuracies associated with conventional two-dimensional bounding box–based methods. Second, 97 quantitative features are extracted by integrating multimodal information from the video. These features include wheelchair kinematics derived from the estimated 3D positions, upper-body skeletal motion obtained via RTMPose, and ball trajectory parameters tracked using TrackNetV3. This representation captures the interdependent dynamics of wheelchair operation, player movement, and ball behavior. Third, a machine learning–based model is constructed to classify stroke effectiveness into three categories: “Win,” “Continue,” and “Lose.” Validation using match data from the Japan Open 2025 shows that, among the classifiers evaluated (including Random Forest and LightGBM), the Random Forest model achieves an accuracy of 78.18%. Feature importance analysis indicates that ball return speed is the feature most strongly associated with stroke effectiveness. The proposed framework enables practical, video-based performance evaluation without wearable sensors.
Download
Paper Nr: 75
Title:

Computer Vision Based Automatic Tracking System for Basketball: Design and Evaluation of a Motorized Camera Prototype
Authors:

Joseph Léonard Kamdem Kamdem and Jérôme Rocheteau
Abstract: Automatic tracking systems are increasingly used in sports to facilitate autonomous video recording and provide data for performance analysis. Commercial solutions already incorporate automatic camera control for team sports, particularly through AI-assisted tracking. However, few studies provide a detailed description and an experimental evaluation of such tracking systems applied to basketball. This article presents the design and evaluation of a vision-based automatic tracking prototype for basketball. It consists of an embedded computing unit, an inference acceleration module, a wide-angle monitoring camera, a motorized recording camera, a motor control interface, and a servomotor. The monitoring camera continuously captures a wide view of the scene that is processed by an object detection model to identify the players and determine the target associated with the current action. The position of this target is converted into an angular correction for the motor control interface, which drives the rotation of the recording camera in order to allow to keep the action near the center of the frame. The article describes the system architecture, the inference pipeline, the target selection strategy, the conversion into angular command, and the gradual control of the servomotor. This study also provides an experimental evaluation of this prototype for autonomous recording of basketball with regard to centering accuracy, tracking stability, target losses, and maintaining the action within the frame. This study also provides an evaluation framework for assessing tracking quality mainly based on the horizontal distance between the corresponding target and the center of the image from the motorized recording camera. Experiments shows that this prototype achieves nearly 90% of target centering.
Download
Paper Nr: 81
Title:

SwimForm: Marker-Free Detection of Head Position Faults in Freestyle Swimming from a Single Above-Water Smartphone Camera
Authors:

Rohan Khandelwal
Abstract: Real-time biomechanical feedback for swimmers has historically required underwater rigs, multi-camera installations, or wearable sensors, limiting access outside elite programs. We present SwimForm, a system that detects head position faults in freestyle swimming from a single above-water smartphone camera, combining a fine-tuned MediaPipe BlazePose backbone with a vector-geometric formulation of head deviation approximately invariant to in-plane camera motion (pan, tilt, roll), flagging faults at 20 degrees from neutral. The model was fine-tuned on 6 swimmers and evaluated on a subject-disjoint pool of 2 additional swimmers (8 total). Under Leave-One-Swimmer-Out cross-validation, reported separately by evaluation unit, the six framelevel folds attain 85.0% mean accuracy (SD 5.4%) and 83.9% recall (SD 6.1%), while the two subject-disjoint clip-level folds attain 84.0% and 88.0% accuracy; pooled over the 50-clip held-out pool, accuracy reaches 86.0% (95% CI 73.8–93.0) and recall 92.0% (95% CI 75.0–97.8). The dominant failure mode is mistaking legitimate breaths for faults (3.7× higher misclassification rate, p = 0.100, not significant at this sample size), motivating stroke-phase-aware deployment. We characterize a systematic shoulder-roll error caused by the swimmer’s own body rotation rather than camera motion; a monocular pseudo-3D world-landmark correction underperformed the 2D formulation at every roll level, attributed to single-camera depth noise in aquatic capture, so we instead introduce a multi-keypoint roll validation signal. Given the small held-out pool and the clustering of frames and clips within individual swimmers, results should be read as a proof-of-concept system and above-water benchmark contribution rather than a validated clinical tool.
Download
Short Papers
Paper Nr: 22
Title:

BreakingBag: A Smart Boxing Serious Game for Enhancing Perceived Motivation and Anticipated Resilience
Authors:

Lucas Haemmerle, Dominik P. Hoelbling, Rene Baranyi, Roland Breiteneder, Karl Pinter and Thomas Grechenig
Abstract: Physical inactivity and obesity are global health challenges, largely driven by a lack of motivation to exercise. Current research suggests that serious games combined with smart sports equipment may offer a partial solution, and the constant rise in popularity of combat sports such as boxing within the fitness domain adds further motivational potential. This work therefore combines and tests both approaches, using novel prototypes of force- and acceleration-measuring smart boxing gloves (“RISE Dynamics Alpha”; RD α) and an Android tablet. Based on expert interviews (N=4), a reaction-based game concept (“Destroy Blocks”) was developed and visually designed to fit the domain. It was evaluated in a field study and survey (N=30) with a well balanced sample of martial arts athletes and non-athletes as well as gamers and non-gamers. Engagement was high: participants rated the overall fun factor 4.80 +/- 0.41 of 5 and motivation to keep playing 4.50 +/- 0.63. Potential routine adherence compared favorably to traditional workouts (3.97 +/- 0.94), though participants criticized an input latency that should be optimized in the future. Based on these initial findings, the prototype was evaluated positively and shows promise for enhancing perceived motivation and anticipated resilience for physical exercise in both performance and recreational sports.
Download
Paper Nr: 26
Title:

Engineering Equity through a Biomechanical and Socioeconomic Revision of Footwear Design for the Female Athlete
Authors:

Maxwell Prempeh and Colin G. Pennington
Abstract: The current landscape of women’s soccer is characterized by a significant disparity between athletic participation and the technological adequacy of essential equipment. Female athletes currently sustain non-contact anterior cruciate ligament (ACL) injuries at rates two to eight times higher than their male counterparts, a phenomenon increasingly attributed to the use of footwear designed for male biomechanics and anatomical structures. This report details the development and empirical validation of the Kɔfa Agilitá, a state-of-the-art soccer cleat engineered specifically for the female foot. Utilizing player feedback from collegiate athletes and rigorous lab performance testing, the project addresses four essential pillars: (1) the reduction of acute and chronic discomfort through a modified foot last; (2) the mitigation of overuse injuries by optimizing rotational traction; (3) the enhancement of performance through psychological confidence; and (4) the improvement of athlete well-being via gender-responsive design. Quantitative analysis demonstrated that the Kɔfa Agilitá successfully reduced peak rotational torque by 64% compared to leading industry competitors. Furthermore, the integration of advanced material science, such as polyethylene-based "Midas Touch" insoles and Clarino Tirrenina suede heel linings, significantly outperformed industry-standard ethylene-vinyl acetate (EVA) foams in durability and compression sets. This research emphasizes that the development of gender-specific footwear is not a luxury but a clinical necessity for the longevity of the female game.
Download
Paper Nr: 28
Title:

Feedback Generation Method for Pitching Instruction Using Motion Analysis Based on Pitching Phase Estimation and Large Language Models
Authors:

Yoshiki Saito, Shimpei Aihara, Yudai Yoshida, Minori Ota, Hiroki Watanabe and Yoshinari Takegawa
Abstract: In baseball pitching instruction, beginners often need guidance that identifies the most important correction point in each phase, but such feedback depends on experienced coaches. This study proposes a monocular-video method that combines pitching phase estimation, skeletal motion comparison, and a large language model. A YOLOv8-based model estimates six pitching phases from annotated images, and two-dimensional skeletal keypoints are interpolated, normalized, and aligned with experienced players’ motions using Dynamic Time Warping. Phase-wise deviations are then calculated as z-scores and summarized for Gemini 3.1, which generates feedback consisting of the primary correction point, the reason correction is required, and an improvement method. The phase estimation model achieved 0.909 precision, 0.884 recall, and a 0.896 F1 score. In an evaluation by seven participants with baseball experience, the proposed feedback achieved an overall appropriateness rate of 0.838.
Download
Paper Nr: 35
Title:

LSTM-Based Badminton Stroke Classification: Comparing OpenPose Keypoint Formats under Event-and-Camera Shift
Authors:

Muhammad Yusuf Fadhlan and Uffe Kock Wiil
Abstract: Badminton is one of the most popular sports globally, yet coaches and sports analysts still rely on conventional methods to evaluate players from match videos. As a first step toward a comprehensive badminton analytics platform, this paper proposes an end-to-end system to classify four badminton strokes (serve, smash, backhand, and lift) using Human Pose Estimation with OpenPose combined with a Long Short-Term Memory (LSTM) classifier. Skeleton keypoints were extracted in both COCO (18 keypoints) and BODY_25 (25 keypoints) formats from broadcast match videos, grouped into 25-frame sequences, and fed into the LSTM model. On same-event test data, both keypoint formats reached 94.73% accuracy, with macro-averaged F1 scores of 94.17% for COCO and 94.03% for BODY_25. Under external validation on footage from a different tournament, recorded with a higher and more distant camera, BODY_25 retained 91.47% accuracy and 90.80% F1, whereas COCO fell to 81.39% accuracy and 79.90% F1. The results are comparable to those reported by prior deep learning approaches, while using only broadcast video without wearable sensors or controlled laboratory conditions. BODY_25 is therefore markedly more robust to this combined event-and-camera shift, with a 3-percentage-point drop compared to over 13 points for COCO, although the study design does not allow the effect of camera placement to be separated from the other differences between the two events.
Download
Paper Nr: 57
Title:

Demonstration of Video Generation Using Generative AI to Support Exercise Movement Acquisition
Authors:

Haruto Tomoda, Yuki Nagano, Chun Xie and Itaru Kitahara
Abstract: Video-based observation is widely used for motor skill learning, broadly categorized into self-feedback and demonstration observation. Self-feedback allows learners to review their movements, aiding in understanding the relationship between body movements and resulting motion. However, ascertaining movement is difficult. By contrast, demonstration observation facilitates understanding of ideal movements but does not always help learners relate the observed motion to their own body movements. To integrate the advantages of these two learning approaches, this study proposes a motor learning support method that uses generative artificial intelligence (AI) to create self-demonstration videos. In these videos, learners appear to perform ideal movements. The method takes a learner’s image and an instructor’s demonstration video as inputs. After applying image preprocessing and pose scaling to account for differences in body proportions, a self-demonstration video is generated using a video generation AI model. Furthermore, to evaluate the effects of each processing step on video quality, a quantitative evaluation based on face similarity was conducted. The results indicated that image preprocessing and pose scaling improved face similarity and more stably preserved the learner’s appearance features. These findings demonstrate the proposed method’s effectiveness in generating high-quality self-demonstration videos.
Download
Paper Nr: 59
Title:

Video-Derived External Load and Technique Metrics for Functional Training Monitoring Using Monocular Smartphone Video
Authors:

Lúcia Sousa, Daniel Canedo, Pedro Santos and António J. R. Neves
Abstract: This paper presents a video-only pipeline for extracting interpretable visual metrics from functional training classes using a smartphone camera fixed for each recording, repositioned between participants and exercises to keep one target participant in frame . Body landmarks were estimated frame-by-frame using MediaPipe, and a Movement Description Language (MDL) with exercise-specific finite-state machines was used to segment repetitions and movement phases. From the segmented signals, the repetition count, set duration, active and rest time, cadence, inter-repetition pause, range-of-motion proxies, and peak angular-velocity proxies are computed. A dataset was recorded with a fixed smartphone camera during six real-world class sessions with a total of twenty-five participants, yielding 550 valid sets and 8331 valid repetitions across 29 exercise types. Against manual annotations, repetition counting achieved an overall mean absolute error of 0.42 repetitions and a root mean squared error of 1.24 repetitions, with 81.8% of sets counted exactly and 89.1% within one repetition of ground truth. Across exercise types, the visual metrics showed coherent profiles, with more explosive movements exhibiting higher cadence and angular-velocity proxies and more controlled movements showing lower cadence and more stable range-of-motion proxies. Round-wise trajectories also revealed participant-specific execution drift across repeated rounds. These results support the use of monocular smartphone video as a low-cost source of external-load and movement-execution metrics for coaching support in functional training classes, while avoiding claims of clinical-grade kinematics or internal-load estimation.
Download
Paper Nr: 72
Title:

A Vignette-Based Testing Protocol for Acute Post-Error Self-Regulation in Tennis: Comparing Cue-Word-Based Self-Talk, Coping-Oriented Mastery Imagery, and Neutral Audio Control
Authors:

Vitus Widmann and Klemens Weigl
Abstract: Evidence from elite tennis indicates that pressure and controllable errors, such as double faults, can increase the risk of subsequent errors, making the immediate post-error window a relevant target for performance-oriented testing. Self-talk and imagery are established mental strategies in sport psychology, yet their translation into a single, brief, audio-guided post-error application remains insufficiently specified. This position paper argues that acute post-error regulation should first be examined through standardised, profile-based testing protocols before stronger claims about applied effectiveness or performance transfer are made. We propose a digitally deliverable protocol comparing cue-word-based self-talk (CWBST), coping-oriented mastery imagery (COMI), and a neutral audio control condition (NAC) after a standardised high-pressure double-fault vignette. The protocol combines pre- and post-assessments of regulation outcomes with process checks and profile-based evaluation. CWBST is expected to show stronger effects on action-regulatory outcomes, such as action clarity and task focus, whereas COMI is expected to show stronger effects on coping-appraisal outcomes, such as confidence, perceived control, and a functional interpretation of tension. The paper outlines methodological contributions, limitations, and directions for future research.
Download
Paper Nr: 74
Title:

Gaze Pattern Acquisition by VR Golf Swing Observation for Instructor Skill Evaluation
Authors:

Kazuma Tanoguchi, Yuto Nakamura, Chun Xie and Itaru Kitahara
Abstract: While sports instruction requires detecting subtle changes in an athlete’s movements and providing appropriate feedback, most research on gaze behavior in sports has primarily focused on athletes. This research investigates the gaze patterns of golf instructors during swing instruction. To reproduce a realistic instructional environment, an observation system is developed that presents 360-degree videos, recorded from front and rear viewpoints using 360-degree cameras, in a virtual reality (VR) environment using a head-mounted display (HMD). The system enables seamless viewpoint switching and playback control, while gaze behavior is measured via built-in eye-tracking. A gaze measurement experiment is conducted with three golf instructors with different levels of instructional experience. After determining the optimal parameters for the Identification by Dispersion Threshold (I-DT) algorithm from the acquired data, the fixation tendencies during the address phase are evaluated. The results show that while the instructors share common fixation areas, such as the hands and clubhead, there are distinct differences in the temporal sequences of their visual searches which are based on their levels of instructional experience. This research thus contributes to the externalization of expert instructional knowledge and the effective transfer of instructional skills to novice instructors.
Download
Paper Nr: 86
Title:

Analysis of the Structure of the Dual Application of IHF Rules 7:11-12 of the Handball Game
Authors:

Igor Gruić and Antonio Butorac
Abstract: The temporal scope of the application of Rule 7:11-12 creates confusion regarding the understanding and acceptance of the rule's application outcomes among all competition participants (players, coaches, and spectators). Specifically, its application during the 0:00–54:59 interval differs from that during the 55:00– 60:00 period, especially due to different application of time-out rules (Rule 2:10, within Clarification 3), psychological responses of referees in tight (draw) results with high expectations from players, coaches and spectators (and their emotional response on fairness on lost/won outcome). The aim of the research was to determine the level of statistical significance of existing differences in the application of IHF handball rules 7:11–12 during different phases of matches ending in a draw. The sample consists of all attacks in which the IHF Rule 7:11–12 (passive play) was initiated and applied during the last 30 matches ending in a draw at major men's competitions (World Championship, European Championship, Olympic Games, Champions League). Indications of differences were identified regarding fouls suffered (a primary characteristic of players' technical-tactical actions) and 2-minute suspensions (primarily regulated by referees based on criteria aligned with the rules of the game). Of greater concern are the "anomalies" in rule application evident at the descriptive level-specifically, instances where the number of passes exceeds four-which, in drawn matches, provide a compelling basis for objections from aggrieved stakeholders.
Download
Paper Nr: 18
Title:

The OEXfit as a Technological Innovation: Applications in Sport-Performance Teaching and Research
Authors:

Colin G. Pennington
Abstract: This position paper explores the multifaceted integration of the OxeFit system, an AI-powered smart gym, as a transformative technological innovation within an Exercise Science Program emphasizing sport performance. Characterized by digital weight resistance, integrated force plate technology, and AI-driven motion analysis, the OxeFit represents a shift toward data-centric methodologies in human performance. This paper details the system’s application across two primary domains: pedagogy and research. In teaching, the OxeFit bridges the gap between theoretical concepts and practical application in biomechanics, exercise physiology, and strength and conditioning. It provides students with real-time visualizations of ground reaction forces, velocity-based training (VBT) metrics, and neuromuscular fatigue patterns. In research, the platform facilitates high-precision investigations into eccentric overload, muscle adaptation, and the efficacy of AI-personalized training protocols. Furthermore, this paper discusses the broader institutional impact, including enhancing student-athlete performance, fostering interdisciplinary partnerships (clinical, athletic, and community), and increasing program recruitment through state-of-the-art facilities. Using OxeFit’s capabilities, academic programs can offer a modernized curriculum that prepares students for the evolving landscape of sports technology while generating diverse avenues for revenue and scholarly output.
Paper Nr: 32
Title:

A Dual-Sensor Wearable System for Real-Time Tennis Motion Monitoring and Analysis
Authors:

Fanhua Ye
Abstract: This study presents a low-cost wearable tennis motion monitoring system based on Arduino UNO, dual MPU6050 sensors, Bluetooth communication, and an Android application. Two MPU6050 sensors were attached to the wrist and trunk to collect real-time acceleration and gyroscope signals during tennis-related movements. The acquired data were transmitted wirelessly to a mobile device for real-time visualization and monitoring. Experimental results demonstrated that the proposed system successfully captured motion variations associated with racket swings, upper-body rotation, and body movement. Distinct signal characteristics were observed between wrist and trunk sensors, while different tennis movements generated distinguishable motion responses. The proposed system provides a practical and portable platform for tennis motion monitoring and demonstrates the feasibility of combining wearable sensors with mobile applications for sports training and movement analysis.
Paper Nr: 39
Title:

Corrective Feedback for Tennis Forehand Stroke Training Using Motion Feature Correlations Analyzed with a Statistical Pose Model
Authors:

Tomoya Shiokawa, Kunio Yamamoto and Masaki Oshita
Abstract: Self-training systems can detect motion features that deviate from expert motions and visualize individual corrections, but independently presented problems may not indicate how to modify the posture as a coordinated movement. We present a corrective feedback method that jointly visualizes a main problem and related problems. Statistical pose models are built from expert tennis forehand key poses, and directional relations between features are estimated by changing one feature and observing the others. At runtime, the system detects problematic features, preferentiallyselects one having another detected related problem as the main problem,and presents related problems using the precomputed relations; thus,the main problem need not have the largest deviation. Two small exploratory experiments used 47 forehand motions from one expert, four novice participants for motion correction, and six participants for subjective evaluation. The results provided preliminary observations consistent with faster reduction of detected problems and easier simultaneous correction; larger studies are needed to assess effectiveness.
Download
Paper Nr: 70
Title:

Bridging the Gap between Data and Coaching Practice in Player Development
Authors:

Mikko V. Pohjola, Tanja Matarma and Elisa Hakamäki
Abstract: Continuous, large-scale athlete monitoring is common in elite sport, including football. Data can support assessments of team performance, individual development, training load, and recovery. However, a holistic view of athlete or team development requires multiple monitoring solutions, extensive data analysis, and translation of findings into coaching decisions. Without dedicated expertise or tools that integrate and filter data for specific coaching needs, monitoring can become a burden rather than an aid. Current solutions do not fully meet the practical demands of varied coaching contexts. A systemic shift is therefore needed from technology-driven development to needs-driven collaboration among data providers and users. This position paper (1) defines invisible monitoring as a decision-oriented approach, (2) describes its current examination in Finnish football, and (3) proposes Knowledge Management as a Service (KMaaS) as a collaborative model linking data production with coaching decisions. Drawing on literature and project materials, including interviews, case studies, and coach questionnaires, it presents KMaaS as a proposition for empirical testing rather than a demonstrated mechanism for improving player development.
Paper Nr: 71
Title:

Predicting Ground Reaction Forces in Volleyball Players' Jumps Using Electromyography and Inertial Measurement Units
Authors:

Somaia Mohamed, Gençer Yarkın, Kerem Altun and Selda Yücel
Abstract: Measuring Ground Reaction Forces (GRF) is essential for evaluating athletic performance and injury risk in volleyball. Physical force plates provide highly accurate GRF data but are confined to the laboratory. This paper introduces a Virtual Force Plate method to estimate continuous GRF waveforms during three different jumps using wearable sensors. We recorded kinematics using Inertial Measurement Units (IMUs) and muscle activity via Electromyography (EMG), from 25 female sub-elite volleyball players. We trained a sequential Convolutional Neural Network and Long Short Term Memory (CNN-LSTM) model to map these signals to GRF outputs, with separate models for each jump type’s landing and propulsion phases. We report the results in terms of RMSE of the force per unit mass of the player, and demonstrate that it is possible to achieve error values in the order of 1 N/kg, particularly for propulsion phase. This offers a practical way to test jump performance without force plates, as well as to monitor volleyball players directly on the court.
Download