Abstracts Track 2025

Area 1 - Computer Systems in Sports
Area 2 - Health and Support Technology
Area 3 - Sport Performance and Support Technology

Area 1 - Computer Systems in Sports

Nr: 65
Title:

How Do Orienteers Know Where They Are? A Web-Based Tool to Study Human Localization Skills
Authors:

Marco L. Della Vedova, Harald F. Franck and Mengyuan Wang
Abstract: How do orienteers determine their position on a map using only visual information and terrain cues? This question lies at the intersection of spatial cognition, sports science, and applied AI, yet it has received limited empirical attention under controlled, data-driven conditions. In this work, we present a novel interactive web application and accompanying dataset designed to systematically study map-based localization in orienteering. Our platform enables researchers, coaches, and athletes to investigate how humans integrate visual information with map reading skills to maintain situational awareness in unfamiliar terrain. Participants are shown short, first-person video clips recorded by orienteers during training. The participant is given the correct starting position on the map before the video begins, and a virtual compass provides orientation during playback. After viewing, the participant is asked to click on the map to indicate their estimated position at the end of the clip. Each interaction records the guessed location, the time taken to respond, and optionally, the user’s confidence. This setup, coupled with eye-tracking technology, yields a rich dataset for analyzing accuracy, timing, uncertainty, and spatial reasoning across different terrain types and user profiles. The tool has dual value: it functions both as a research instrument to study human localization behavior, and as a training and evaluation platform for orienteers at different skill levels. It supports controlled experiments while maintaining ecological validity, since the visual input comes from real outdoor runs rather than synthetic environments. The contribution of this work is threefold: 1. We introduce a new dataset of human localization attempts in orienteering, aligned with video data and high-resolution maps. 2. We provide an open, modular platform to run controlled localization tasks online, enabling reproducible experiments and comparisons between individuals or groups. 3. We report insights from pilot studies with participants of varying skill levels. Early results suggest that elite orienteers tend to respond faster and more accurately, particularly in open or semi-open terrain. In contrast, heavily forested or visually ambiguous areas lead to decreased accuracy and increased uncertainty across all groups. The system is designed to be easily extendable: it can be used to investigate questions about decision-making under uncertainty, spatial learning over repeated trials, or even cross-cultural differences in navigation strategies. It also offers potential as a benchmarking tool for evaluating AI-based localization systems that aim to mimic human reasoning in complex environments.
Nr: 68
Title:

Tactical Support System for Mölkky Using Physics Simulation and PPO-Optimized Evaluation Function
Authors:

Sora Matsubara and Masataka Imura
Abstract: Mölkky is a sport in which players take turns throwing a wooden pin called “Mölkky” underhand at a set of numbered wooden pins called “skittles” to score points. The winning condition is to be the first player to reach exactly 50 points; therefore, tactical decision-making is essential. Players utilize four different throwing techniques to selectively hit high-scoring and easy-to-hit targets based on the game situation. However, this optimal target selection depends on players' experience and intuition, and no quantitative guidelines are available. This study proposes a new tactical support system integrating computer vision, physics simulation, and deep reinforcement learning to support this complex decision-making process. The proposed system consists of three interconnected modules. First, the Field State Acquisition module determines the positions of all skittles, either automatically via a Python-based image processing pipeline using homography transformation or by manually entering each skittle's position. The skittle positions are transmitted to the physics simulation on Unity via socket communication. Second, the Physics Simulation module, using the Unity engine, simulates the physical result of a throw to any landing point, predicting the immediate score and subsequent skittle configuration. Third, the Tactical Evaluation and PPO-based Optimization module defines a flexible evaluation function to calculate the strategic value of a game state. This function is structured as a weighted sum of key factors, including potential score, various risk factors (such as skittle density and throwing distance), and the estimated probability of a successful throw. The throw success probability is modeled using a decay function dependent on throwing distance, whose parameters—a difficulty coefficient and a decay constant—were determined through prior offline experiments. The primary challenge lies in determining the optimal weights for the scoring and risk factors. To solve this, we employ a framework to train agents using Proximal Policy Optimization (PPO). The optimal weights are determined by identifying the configuration of the agent that achieved the highest win rate after training through self-play. A functional prototype integrating the three modules was developed. The trained PPO agent learned a sophisticated policy that prioritizes strategic advantage over simple score maximization in specific scenarios. The system generates a heatmap to visualize the tactical value calculated by the evaluation function, providing a data-driven assessment that allows players to evaluate both reward and risk intuitively. This study presented a framework for advanced tactical analysis in Mölkky by combining computer vision, physics simulation, and PPO reinforcement learning. This approach effectively captures and optimizes complex strategies in target-based sports. Future work will focus on validating the learned policies against expert human players.
Nr: 72
Title:

Estimating Space for Football Match Using Dominance Distribution Model
Authors:

Masataka Imura and Ryo Yoshimura
Abstract: In football, space refers to the area of the pitch that can be used for attacking purposes. Space is an important factor in tactical analysis because goals are often scored by using space effectively. However, during a match when players and the ball are constantly moving, finding space is difficult for both the players on the pitch and spectators with a bird's-eye view. This study proposes a method of quantitatively estimating the space in a football match based on the positional relationship between players and the ball, player speed, and players’ offensive and defensive roles. The proposed method involves estimating a dominance distribution field representing which team dominates each position on the pitch, and then estimating space from the dominance distribution. The positions and velocities of the players and the ball are obtained from oblique bird's-eye view images of actual matches in which the entire pitch is visible. This is achieved using a machine-learning model. To obtain the positions of the players on the pitch coordinate, the oblique bird's-eye view image of the actual match is converted into a true overlooking view image. This transformation is performed using homography. The proposed method assumes that dominance on the pitch is determined by the situation of the players. A player's situation is represented by position, speed, distance to the ball, and whether the player is an attacker or defender. A player's dominance is higher at the closer area to the player and is biased according to the position, speed, and relative position to the ball. Therefore, a player's dominance is represented by a two-dimensional multivariate normal distribution. The players' situation is reflected in the center and the variance-covariance matrix of the multivariate normal distribution, according to the following criteria: (1) The farther away from the ball, the wider the dominant range; (2) The dominant range is wider in the direction of movement of player and narrower in the perpendicular direction. The center of the dominant range is displaced in the direction of movement of player; (3) Dominance increases in the rear area directly opposite the ball. These criteria are represented by the parameters in the model. The parameters are optimized so that the dominant distribution field estimated by the proposed method matches that subjectively indicated by experienced footballers. The dominance distribution of each team is calculated by adding up the dominance distribution of the players, and the areas of space are estimated based on the difference between the dominance distributions of the teams.

Area 2 - Health and Support Technology

Nr: 26
Title:

Heart Rate and Physical Activity as Indicators of Mental Load in Youth Golf: Insights from Wearable Monitoring in Training and Competition
Authors:

Jasper Gielen, Elise Heirman, Katrien Buysse, Romain Meeusen and Jean-Marie Aerts
Abstract: Golf performance relies heavily on cognitive factors such as focus, arousal, and routine, particularly during putting. While lab-based tools like EEG and eye-tracking have advanced our understanding of mental load in precision sports, they remain impractical for real-world use. Our work explores how wearable technology, specifically heart rate (HR) and accelerometer-based physical activity (PA), can be leveraged to assess mental load during golf in training conditions as well as actual competitive gameplay. Seven youth golfers (ages 13–16, hcp index < 12) participated in a field-based putting exercise, attempting to score five consecutive 3-meter putts while wearing a validated chest strap for HR and PA monitoring (Zephyr Bioharness 3.0). Moreover, the same data are currently being collected during golf competitions. HR and PA were recorded at 1 Hz, and pre-shot values were extracted to reflect the athletes’ physiology. These data were matched to putting outcomes via video annotations. Two core questions guided our analysis of the data collected during training: (1) Do HR and PA levels differ prior to made vs. missed putts? (2) Does the result of a putt influence the physiological state leading into the next attempt? Our analyses related to the first research question showed no consistent predictive value of pre-shot HR or PA for immediate performance. However, in line with the second research question, it was observed that pre-shot HR increased after a successful putt and decreased after a miss. In contrast, PA decreased after scoring, indicating that cardiovascular changes were unlikely due to physical exertion. These patterns suggest that wearable-derived HR data may reflect shifts in cognitive demand, offering a window into mental load of the involved youth golfers. Importantly, the study demonstrates the feasibility of integrating wearable monitoring into training environments without disrupting natural routines. This approach enables athletes and coaches to reflect on training responses objectively, using physiological feedback to frame conversations about pressure and focus. The use of wearable technology thus bridges sport-specific insights with scalable digital methods for athlete monitoring. To build on these findings, we are currently collecting similar data during competitive rounds of golf for the same athletes, where environmental demands, performance stakes, and pacing differ markedly from training. By examining HR and PA patterns in tournament settings, we aim to test whether the same post-performance dynamics hold and whether competition-specific stressors induce distinguishable physiological profiles. These ongoing measurements will be included in our conference presentation if accepted. Our contribution to the icSPORTS conference underscores the potential of wearable technology to assess mental load in real-world sports contexts. By merging physiological sensing with applied performance analysis, we offer a practical framework for understanding how young athletes physiologically adapt to varying performance pressures, both in training and in competition. Please note that the work on physiological monitoring during putting training is currently under review for publication in a journal. We attach the latest version of the manuscript to this submission.
Nr: 64
Title:

Sensory and Perceptual Function in Elite, Fast-Paced Ball-Sport Athletes: A Systematic Review
Authors:

Jake Gerard Tiernan, Sarah Lilly Fitzpatrick, Catherine Fassbender, Redmond O'Connell and David Patrick McGovern
Abstract: Elite, fast-paced ball sports place significant pressure on athletes to quickly and accurately process information in their environment, in order to make the right decisions on the field of play. While elite athletes are traditionally evaluated based on the outcomes of their decisions (e.g. goals scored, passes completed), the psychological and physiological mechanisms by which these athletes process information in order to make these decisions are of crucial importance. Specifically, sensory processing, which describes how information from the environment is obtained through the senses such as vision and hearing, and perceptual processing, which describes how sensory information is organised and interpreted to create meaningful representations of the environment, are two stages of the information processing cycle which are thought to play a key role in elite sports performance. As such, a significant amount of research has been conducted looking at the sensory and perceptual skills at play in elite sport. However, the research on elite athletes specifically has yet to be systematically reviewed. Therefore, the objective of this review was to explore the characteristics that define the sensory perception of elite athletes participating in fast-paced ball sports. Specifically, it aimed to determine whether these athletes possess enhanced sensory perception as compared to both non-athletes and sub-elite athletes, and whether sensory perception is correlated with sport performance metrics. This systematic review was conducted in line with the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) framework (Moher et al., 2009), and was pre-registered on the Open Science Framework website (project: https://osf.io/wnyh8/, pre-registration: https://osf.io/krfya). This review included only studies which examined sensory and/or perceptual function via domain-general tasks, devoid of any sport-related information. Furthermore, this review included only studies that investigated elite athletes, as defined by the participant classification framework (McKay et al., 2022). PubMed, Web of Science and Scopus were searched for relevant studies. A narrative synthesis was employed to describe the findings of included studies. 18 studies met the inclusion criteria and were reviewed. The findings from these studies demonstrate partial evidence for enhanced visual function in elite athletes as compared to non-athletes, but not athletes from different levels of competition. This review also finds evidence of differences in visual function between athletes of different sports, and between athletes of different positions within the same sport. Furthermore, significant correlations were noted between visual function and baseball batting metrics. The findings from this review highlight that while enhancements in visual function may be a prerequisite for sports participation, they may not reliably distinguish athletes from different skill levels. Nevertheless, these findings have implications for the design of sports vision training strategies, which should account for the individual, sport-specific demands of each athlete. Future research should aim to explore perceptual function in greater detail, which may act as an important intermediary between low-level visual skills and higher-level perceptual-cognitive skills which play an important role in elite sport performance.
Nr: 70
Title:

Analysis of Muscle Activity During ASLR with a Knee Extension Brace: Potential Implications for Rehabilitation in Patients with AMI
Authors:

Mateusz Kamiński, Andrzej Jacek Frankiewicz, Gabriel Gipsiak, Tomasz Satławski and Justyna Kędziorek
Abstract: Introduction: Arthrogenic muscle inhibition (AMI) is a significant clinical issue often posing challenges for physiotherapists. The Active Straight Leg Raise (ASLR) exercise, considered safe and frequently implemented during early-stage rehabilitation following knee surgery, may be difficult or impossible for patients with AMI. Findings suggest that muscle activity when using the brace is lower compared to that during resistance or isometric exercises but remains present. Objective: The purpose of the study was to evaluate quadriceps muscle activity during ASLR with the limb supported by a knee extension brace. Additionally, this study aims to assess the potential role of a knee extension brace as an adjunctive physiotherapy tool in managing AMI. Materials and methods: A prospective study was conducted on 39 healthy participants (age: 23.8 ± 3.4 years; BMI: 23.5 ± 2.5 kg/m²). Following a standardized warm-up, maximal voluntary isometric contractions (MVIC) of the quadriceps were measured in two positions: seated (90 degrees knee flexion - MVIC90) and supine (full extension - MVIC0). Performed four ASLR variants: no brace (p1), brace (p2), brace + voluntary extension contraction (p3) and brace + relaxation (p4). sEMG was recorded using a wireless Trigno Research System (Delsys, USA). Electrodes were placed according to SENIAM guidelines on rectus femoris (RF), vastus medialis oblique (VMO), and vastus lateralis oblique (VLO). Results: The brace did not affect RF (p = 0.322, 95%CI [-0.156, 0.475]) and VMO activity (p > 0.05, δ 0.05, 95%CI [-0.21, 0.30]), but significantly reduced VLO activity (p = 0.026, δ 0.18, 95%CI [-0.08, 0.42]). Muscle activation ranked as follows: p3>p1>p2>p4. These rankings achieved significant differences in all comparisons excluding p1-p2 which remained non-significant in 2 (RF, VMO) out of 3 muscles. Conclusions: The knee extension brace reduces VLO activity without affecting RF and VMO during ASLR. Ranking muscle activation across trials allows precise progression and regression of the ASLR, potentially serving as an effective component in rehabilitation programs for AMI patients. Acknowledgements: ClinicalTrials.gov ID: NCT06821477, approved by Senate Commission of Research Studies Ethics (SKE.0030.12.2025).

Area 3 - Sport Performance and Support Technology

Nr: 56
Title:

Comparison of Load-Velocity Profile in Unilateral Upper-Limb-Loss and Non-Impaired Swimmers at Matched Performance
Authors:

Adrián Febles-Castro, Jesús J. Ruiz-Navarro, Óscar López-Belmonte, Ana Gay, Julia Oliva-Álvarez and Raúl Arellano
Abstract: The load-velocity profile has gained attention as a method for assessing specific mechanical sprinting capacities in non-impaired swimmers. However, its application in Paralympic swimming remains largely unexplored. Investigating these profiles may help to understand how asymmetries affect mechanical responses to resistive load in para-swimmers. This study aimed to explore the mechanical differences in load-velocity profiles between two swimmers with equivalent competitive performance: one with total unilateral upper-limb-loss (S8 class, Paralympic World-Championship finalist) and one non-impaired (national-level). Table 1 presents their anthropometric and performance characteristics. Each swimmer performed a load-velocity profile test, consisting of three 20m front-crawl semi-tethered push-off sprint trials with 6-minute rest between each, using a 1080 Sprint 2 device (standardised warm-up). The loads used were 1, 3, and 5kg for the para-swimmer and 1, 5, and 7kg for the non-impaired swimmer. The para-swimmer's external loads were adjusted based on a pilot testing to ensure comparable relative effort between participants, avoiding excessive sprint duration and disruption of technique. A linear regression was fitted to each swimmer's load–velocity data by plotting average horizontal velocity (10-15m section) against the applied load. The theoretical maximum velocity (V$_0$) and load (L$_0$) were calculated using the intercept of the regression line with the vertical and horizontal axes, respectively. The slope was calculated as slope=–V$_0$/L$_0$. The para-swimmer showed lower V$_0$ (1.47m·s$^1$) and L$_0$ (10.13kg), and a steeper slope (–0.14m/s·kg$^1$) compared to the non-impaired swimmer (V$_0$=1.86m·s$^1$; L$_0$=18.35kg; slope=–0.10m/s·kg$^1$)(Figure 1). Despite matching competitive performance, the para-swimmer showed a more pronounced decline, with a greater velocity drop in response to increased load. In contrast, the non-impaired swimmer maintained a higher maximal velocity and was able to displace heavier loads, showing greater propulsion capacity. This reflects the limitations in propulsive force production in the para-swimmer. Swimmers with unilateral propulsion lack the opportunity for contralateral force generation during the non-propulsive phase of the stroke. This may result in lower minimum force values compared to non-impaired swimmers. The reduced ability to maintain continuous propulsion may further hinder the capacity to overcome same external load. These observations support the notion that the unique locomotor pattern and asymmetry inherent to the para-swimmer may lead to distinct mechanical responses under external loads. Load-velocity profile reveals functional differences that may not be detected from performance times solely. These findings highlight the potential of this approach to identify individual mechanical responses in para-swimmers, contributing to a better understanding of how asymmetries influence loaded sprinting performance. This could provide valuable information for the application of specific testing protocols and training strategies scaled to defined impairment profiles. This study was approved by the University of Granada ethics committee (CODE: 3256/CEIH/2023) and it is part of the national project PID2022.142147NB.I00 (SWIM III) funded by MICIU/AEI/10.13039/501100011033/ and “ERDF A way of making Europe”, by the “European Union NextGenerationEU/PRTR”. AFC holds an FPI fellowship which is funded through the mentioned grant.
Nr: 67
Title:

Development of Machine Learning Models for Prediction of Instructional Voices for Decision Training in Football
Authors:

Kodai Ishikawa and Masataka Imura
Abstract: With the advancement of VR technology, a number of systems have been developed to train cognitive and decision-making abilities in football from a first-person perspective. However, the voices of players giving instructions, which are important for situational judgment, have not been incorporated into these systems. These instruction voices are important elements in football, as the voices help teammates recognize the situation and make decisions. The reason why these voices have not been incorporated is that there are no systems to determine which players call out instruction voice in which situations. In this study, to enhance the practicality of the football situational judgment training system, we will construct a machine learning model that can predict the moment when instruction voice is called out and the player who calls out the instruction voice, based on match information such as ball's position, players' positions, and orientations. In this study, the instruction voice "Look ahead" was targeted. We used match information of 10 games obtained from the RoboCup Soccer Simulation League. Since the match information of the RoboCup Soccer League contains no instruction voices, we developed a survey system for football experts to identify the situation in which the instruction voices were called out. The situations are characterized by the frame number, the player who called out the instruction voice, and the player who received the voice. Similar situations within ±5frames were also included as the situation when the instruction voices were called out. The inclusion criteria were that the direction of the ball's movement did not change, the total movement distance of all players was within a certain range, and the orientation of the player receiving the instruction voice remained constant. To predict both the moment when the instruction voice was called out and the player who called out the instruction voice with a single machine learning model at the same time was difficult. Therefore, the proposed system divided the prediction into two machine learning models: one to determine the moment when the instruction voice was called out and another to identify the player who called out the instruction voice. To train these machine learning models, we tested the following machine learning algorithms: SVM, Logistic Regression, Random Forest, XGBoost, LightGBM, and AdaBoost. For the machine learning model to determine the moment, the following features were adopted as inputs: ball position, player positions and orientation, and the weight of each player based on the distance between the ball and the player. The aforementioned features from the previous frame were also used. All positions and orientations were converted to the ball center coordinate system. As a result, LightGBM achieved the highest accuracy, with a recall rate of 0.55 and a precision rate of 0.71. For the machine learning model to identify the player, the following features were adopted as inputs: ball position, player positions, and orientation. All positions and orientations were converted to the center coordinate system of 10 players from one team (excluding the goalkeeper). As a result, XGBoost achieved the highest accuracy, with a recall rate of 0.62 and a precision rate of 0.55. These results revealed that the proposed method can predict instruction voice with a certain degree of accuracy.