In conventional research on Karate technique recognition and improvement point detection, it has not been possible to indicate which parts of an incomplete technique need improvement. In this study, to detect areas for improvement, we defined the motion of techniques using three action features: trajectory, left-right variation, and up-down variation. For each action feature, we trained a recognition model using LSTM and developed a method for technique recognition and improvement point detection. For 440 instances of 11 types of correct techniques, the model achieved a 98.9% recognition accuracy. For 193 instances of 11 types of techniques including incomplete ones, it correctly indicated improvement points with an accuracy of 70.9%.

When instructing others in sports movements, evaluations and communications are often qualitative, with quantitative assessments limited to localized angles, positions, and velocities. In this research, we evaluate differences in movements between sports experts and beginners using motion curved surfaces created by uniquely connecting human joint positions and time series using our previous method. These motion curved surfaces represent velocity, acceleration, and curvature using color gradients. We can represent rhythm and timing by analyzing the differences in velocity and acceleration generation on the behavioral surfaces of experts and beginners. In addition, the shape and curvature of the behavioral surfaces simultaneously illustrate the movement trajectories. Sports movements are tracked using Microsoft Kinect to capture full-body joint positions. After that, the full-body movement trajectories are converted into behavioral surfaces to evaluate the skill of primarily upper-body sports movements, such as the table tennis forehand stroke and etc. Differences in movements between novices and experts are quantitatively clarified from the shape of the behavioral surfaces and the patterns of changes in velocity and acceleration.

In recent years, objective analysis based on data is required in the field of sports. However, commercially available motion capture and force plates are expensive and require large-scale facilities in most cases. In this study, we propose a new motion analysis system that integrates GRF (Ground Reaction Force) using an inexpensive pressure sensor and markerless motion analysis using a pose estimation model (MoveNet) that can be executed on a mobile terminal, for the purpose of performance improvement and injury prevention of baseball motion. We propose a new motion analysis system to improve the performance of baseball motion and to prevent injury. In the experiment, we developed a form analysis application for mobile terminals that can estimate and visualize the load of the axis foot and the stepping foot using the vertical floor reaction force obtained from a simple force plate with a pressure sensor and the pitching motion extracted from the pose estimation model. As a result, it was confirmed that the magnitude and direction of the floor reaction force could be estimated from the video, and the baseball motion could be evaluated and analyzed.

The purpose of this study is to elucidate the intersegmental relationship of body segments during running movement, with the aim of applying the findings to sports instruction and performance enhancement. As a preliminary step towards understanding running movement, this report focuses on walking movement. Given that both walking and running are periodic movements, translational accelerations in the vertical and anteroposterior directions are analyzed in the frequency domain. Specifically, the relationship between the lower trunk and lower limb segments (thigh and shank) is examined by comparing the amplitude and phase of each frequency component between the two segments at various walking cadences. As a result, both common and distinct patterns in the amplitude and phase differences between the lower trunk and the thigh and shank were identified. These findings of this study suggest that the thigh and shank move in relation to the lower trunk during walking, thereby contributing to trunk stability and the fine-tuning of movement.

Modeling and analyzing matches are challenging in sports due to the continuous interactions among multiple players. Rule-based models struggle to capture all players and scenarios, while reinforcement learning models often lack interpretability in evaluation values. This study explores whether an expandable rule-based state variable reinforcement learning model using spatiotemporal data can effectively analyze soccer tactics. We defined state variables and actions based on interviews with a head coach from the Kanto Soccer Division 1 League. All players and scenarios were modeled, reinforcement learning was conducted, and state-action values （Q-values） were calculated. Results revealed a trade-off between action prediction accuracy and TD error. Although no correlation was found with traditional metrics such as xG or goals, the model effectively evaluated high-risk tactics like counterattacks and defensive breakthroughs compared to the baseline. Furthermore, visualizing Q-values and state variables enabled the interpretation of Q-values and individual players' evaluations. This approach shows promise for enhancing coaching strategies and player assessment.

We developed an application to quantify performance of routine practices in trampoline gymnastics．We applied deep learning-based estimation technique to obtain posture information．General pretrained model has lack of recognition accuracy for trampoline video．The accuracy was reached over 90 % by adding about 8,500 images to train for our model．posture data from deep learning could be used to calculate information related to scoring and coaching．We developed a web application that implement these technologies for player's practice．Our cloud application is available by tablet device and the internet without GPU．The effect of using this system are observed to make it easier for the coaches to explain to the players and grasp the characteristics of their performance．Future studies are needed performance, cost of cloud application and support variety kind of videos.

The forward-step and false-step are common techniques to start sprinting in various directions from a standing position. The aim of this study was to provide a new insight for clarifying the step technique to make the subsequent sprinting direction less predictable. The dataset consisted of the motion capture data and the ground reaction force data. These data were recorded in the sprint start in rightward and backward directions form a standing position by using either the forward-step or the false-step. A recurrent neural network that outputs the subsequent sprinting direction from the kinematic data was constructed by supervised learning with the data of seven subjects. The data of the other three subjects were utilized to test the prediction by the RNN, and the time-series data of accuracy was calculated. The function of findchangepts (MATLAB 2024b) was used to detect abrupt changes in the time-series data. The abrupt changes were found in the frames of 50^th and 111^th in the forward-step and 38^th and 93^rd in the false-step from the movement initiation. These results indicate that the accuracy of the RNN in predicting the subsequent sprinting direction increase abruptly earlier in the false-step than the forward-step. It was suggested that the forward step can make the subsequent sprinting direction less predictable than the false step.

Sport climbing has gained in popularity in recent years, particularly with its appearance as an Olympic sport. Among its disciplines, bouldering involves short, high intensity climbs that require diverse movement strategies. While motion analysis in climbing has been explored, most studies focus on beginner instruction and use coarse phase-based segmentation. Few have addressed the concept of specific "moves"—distinct climbing actions commonly used in practice and recognized by climbers as key to performance. This study proposes a method to detect and classify climbing moves from pose estimation data as a step toward performance quantification. We constructed a dataset of 71 fixed-camera climbing videos and extracted frame-wise skeletal sequences using a fine-tuned 2D pose estimation model. Based on these sequences, we trained a Transformer-based classification model to assign one of seven predefined move labels to each frame. Additionally, a 5-class setting was evaluated by merging infrequent labels. The results demonstrate the potential for move-level classification using pose data alone. Future work will focus on real-world deployment and integration with performance assessment tools.

In Japan, new information provision for the enjoyment of sports spectating. In the marathon race, RFID tags are attached to athletes and receivers are installed at specific points to measure their passing times, and this information is used in marathon broadcasts to provide viewers with lap times and predicted finishing times, in addition to commentary on the race development. However, because it is difficult to measure performance information such as the speed, pitch, and stride of each athlete, this information has not yet been provided. In cooperation with Kansai Telecasting Corporation, the authors have developed a technology for detecting and tracking players and estimating their pitch and stride using deep learning from live TV images. However, issues remained in real-time processing and manual operation. Therefore, this research, existing technology was improved and its applicability to real environments was verified. This technology has the potential to revolutionize sports broadcasting by providing a new viewing experience for marathon broadcasts.

The objective of this study was to investigate coordination between the lower and upper limbs during the golf swing by estimating muscle exertion based on inverse dynamics using a musculoskeletal model. The golf swing was represented by obtaining positional coordinate data from markers attached to the subject's body. These markers were acquired using a 3D motion capture system. The data were then entered into a musculoskeletal model scaled to the subject's physique. The motion of the club was also represented by introducing the positional coordinate data from markers attached to the grip, shaft and head into the club model. The inverse analysis was employed to estimate muscular exertion in the upper and lower limbs. Separately, the muscle parts believed to contribute significantly to the swing motion were extracted for each of the left and right upper and lower limbs. The coordination from the lower limbs to the upper limbs in the swing motion was investigated based on results of the muscle exertion. The time histories of the muscle exertion of the extracted muscles exhibited a tendency to be distributed to the left and right sides, in addition to being transmitted from the lower limbs to the upper limbs, in order to perform the required motion in each phase of the swing.

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 proposed a method for quantifying motion characteristics by analyzing the network of whole-body coordination based on the estimation of coordination using inter-region movement entropy. Specifically, we proposed a method for quantifying the number of edges connecting joints and the probability of occurrence of edge generation based on multiple coordination networks calculated continuously from a series of forms. Furthermore, we applied the proposed method to a prototype system and evaluated the batting forms of actual amateur baseball players. As a result, we quantified that the occurrence of edges 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.

This study experimentally clarifies the relationship between agonist and antagonist external oblique muscle activity, torsional movement, and the transferability of torso rotation. Additionally, we visualized torso rotation, muscle activity, and transferability to examine the effects of visual interventions. The analysis revealed that the timing of peak muscle activity played a dominant role in transferability. Specifically, earlier peak activation of the antagonist and later peak activation of the agonist muscles were associated with improved transferability. Moreover, the size and velocity of torsional movement were also important factors. These findings suggest that optimizing the timing of agonist and antagonist muscle activity, as well as torsional movement, can enhance the transferability of torso rotation. Developing an intervention system based on these findings may contribute to improved athletic performance and injury prevention.

This study aims to quantitatively identify individual differences among curling stones. Using a delivery robot and a stone tracking system, the researchers collected and analyzed 899 shot position data points from 10 stones. The experiment was designed to control for human error and ice condition variability, dividing trials into three sequential groups to observe changes over time. Results showed that vertical displacement (Y-coordinate) varied significantly during early trials (1–3 shots), likely due to unstable ice surface conditions. In contrast, horizontal displacement (X-coordinate) was generally small and stable. ANOVA and Tukey's HSD tests revealed that certain stones had distinct movement patterns: R1, R5, and R6 differed in both of X and Y directions; Y1 and Y7 in the Y direction only; and Y5 and Y6 in the X direction only. These findings provide quantitative evidence supporting the subjective notion of "inconsistent" stones in curling. Future work will involve accumulating more data and comparing it with athletes’ perceptions to develop objective indicators of stone behavior. The outcomes of this research may also contribute to training support, stone quality control, and ice surface evaluation.

This study quantitatively evaluated the stability and reproducibility of the delivery motion in curling by analyzing plantar pressure data. Participants were categorized into beginner，intermediate，and advanced groups，and their delivery motions were divided into four phases for detailed comparison of pressure changes. Using a foot pressure distribution measurement system，data were collected from nine participants. The results showed that beginners exhibited low reproducibility in both the X and Y axes of pressure balance，indicating unstable movements. In contrast，intermediate and advanced players demonstrated higher reproducibility，with advanced players showing particularly stable pressure balance along the X-axis. This study provides quantitative insight into the relationship between skill level and foot pressure during delivery.

In the sport of curling, understanding the causes of missed shots is critical for improving performance and refining strategies. This study aims to design and implement a system to support the analysis of missed shots in curling. Building on the existing digital scorebook system iCE, we designed a method that integrates match records and video footage to visualize when and under what conditions missed shots occur. We also defineed missed shots used a shot score, and classified their causes to analyze patterns within matches. The designed system enables users to filter shots by various conditions and replay corresponding video clips, supporting athletes and coaches to efficiently review and reflect on specific situations. In the future, we aim to implement our design as a new functions of iCE system and conduct experiments to evaluate the perfprmacnce with expert curlers and coaches.

In this study, I analyze a dataset comprising approximately 150 pitching metrics for 350 pitchers from Nippon Professional Baseball (NPB), as provided by DELTA Inc. These metrics encompass a wide range of performance indicators and are categorized into nine distinct groups. I conduct principal component analysis (PCA) to reduce dimensionality and identify key underlying factors that explain the variance across these metrics. Based on the results of the PCA, I perform a clustering analysis using a Gaussian mixture model (GMM) to classify pitchers into multiple distinct groups. This analytical framework enables us to uncover latent structures within the data and to evaluate player similarity from a multidimensional perspective. The resulting classification not only facilitates more precise comparisons of player profiles but also allows for strategic insights into player substitutability and trade planning. my approach offers a novel methodology for player evaluation that integrates advanced statistical techniques with a comprehensive set of performance indicators. By providing a data-driven foundation for roster management and talent acquisition, this research contributes to the evolving landscape of analytical decision-making in professional baseball.

In recent years, the use of ICT has become popular in the field of sports. A technology for quantitatively analyze performance based on data on player movement and team strategy has been developed. In basketball, data-based performance evaluation has been attracting attention. However, the process leading to the decision to shot and the factors involved in success or false have not been adequately analyzed. Therefore, this research, we classify the success or false of a shot based on the player's positional information and analyze the factors. Eventually, it is hoped that these findings will contribute to improved team tactics and coaching policies, leading to improved athletic performance.

In this research, we examined the degree to which pressing, an element strongly required for defenders in one-on-one (duel) situations in football, contributes to the outcome of duel, and how the pressing should be done. The research consisted of 320 duels played by nine selected sixth-grade elementary school players from the academy of a Japanese top league football club. It was found that while some players pressed hard against the attacker, resulting in a low winning percentage, those players with a high winning percentage generally tilted their bodies backward early in the pressing. This study showed that not only the degree of pressing, but also how early the defender tilted backward and started preparing to face the attacker can contribute to the outcome of the duel.

This study proposes a performance metric for evaluating tennis serves using kernel density estimation (KDE) applied to data collected via the Hawk-Eye system. Focusing on first serves on the deuce court from 377 men’s singles matches at the Australian Open between 2022 and 2024, the study analyzed 12,463 valid service points from right-handed players. Only right-handed players and valid first serves were included. KDE was used to model the spatial density distributions of serve landing locations and ball speeds that led to service winners or aces. High-density areas represent effective service zones associated with point acquisition. The bandwidth and other KDE parameters were optimized using grid search. To validate the model, logistic regression was conducted on 2024 data to assess the relationship between estimated density and point success. The results revealed that higher serve speeds tended to be effective when directed toward the center line, while lower speeds were more successful when targeting the wide area. The model demonstrated strong predictive performance(Accuracy = 0.76，Precision = 0.68，Recall = 0.60，F1 Score = 0.64, AUC = 0.81), and an increase of 0.1 in the estimated density was linked to a 2.764 times higher chance of winning the point. These findings suggest that KDE-based density modeling enables quantitative assessment of serve strategies and provides valuable feedback for coaching by visualizing effective serve zones based on empirical match data.

When the data entered by badminton analysts, which involves subjective judgement, was examined using intraclass correlation coefficients, a very high degree of consistency was found. Increasing the amount of data to be collected and continuing discussions on the criteria for determining aces and errors is expected to lead to the development and construction of an automatic input system using AI in the future.

Negative Magnus lift force acts on a sphere or cylinder rotating with backspin within the critical regime at certain rotational speeds. It was previously believed that negative lift could not occur on a baseball with backspin, such as a forkball or splitter, due to the influence of the ball’s seams. However, recent computational fluid dynamics (CFD) simulations have revealed that while a forkball experiences time-averaged positive lift due to backspin, unsteady negative lift can emerge depending on the rotational speed, seam orientation, and stitch position⁽¹⁾. To conduct experiments across a wide range of finely controlled rotational speeds, we developed an experimental apparatus capable of accurately adjusting the rotational speed of a sphere or baseball. We then carried out wind tunnel tests to investigate the conditions under which forkballs and splitters exhibit sharp drops. Our results indicate that the maximum negative value of time-averaged lift occurs at a wind speed of 144 km/h and a rotational speed of approximately 545 rpm.

The pre-release dynamics determine the rotation of a baseball pitch at the moment of release. The variable-length pendulum model, which is derived from the center of curvature of the ball’s trajectory and the point of force application, is utilized to clarify the relationship between the mechanism governing the ball’s rotation and the efficiency with which the force is applied to the ball.

In sports movements, there is a phenomenon in which velocities are sequentially added from the center to the end of the body. If the velocity information of one joint can be used to predict the velocity information of a distal joint, it is expected that motion analysis can be performed more easily and at a lower cost. The purpose of this study is to construct a method to estimate the velocity of the upper limb joints in baseball batting using the Fibonacci sequence. The velocities of joints in the left upper limb were estimated by a Fibonacci sequence model with the velocity of the anterior superior iliac spine as the first term. It was compared with those measured by motion capture for the tee-batting condition of four experienced baseball players. The intraclass correlation coefficients between motion capture and Fibonacci sequence model were high (ICC(3,1) = 0.87). The proposed method achieved a root mean square of the error of 1.68 m/s for velocity, which is less than half of the error of the geometric sequence model. These results suggest the possibility of estimating the velocity of distal joints from a single joint in the batting motion by the Fibonacci sequence model. It is expected to be a basic algorithm that will contribute to the establishment of a simple motion analysis using a small number of sensors and low-cost cameras.

The purpose of this study was to investigate swing skills in baseball batting using dynamic simulation and dynamic contribution analysis methods. To replicate time constraints, a badminton shuttlecock launching machine was used for batting practice, focusing on trials involving inside pitches. The bat and upper body were modeled as rigid body links, and the equations of motion for the system were derived. Using these equations, dynamic contribution analysis was applied to the motion of the center pitch trial. It was found that the acquisition of velocity in the direction of the bat's long axis was contributed by the internal and external torque of the right shoulder, the flexion and extension torque of the left elbow, and the left and right rotational torque of the torso. Based on this information, the joint torques were adjusted during the dynamics simulation. The results indicated that, compared to the impact position of the original swing, the simulation swing brought the distance from the sweet spot closer by an average of 3.6 cm, and additionally, it approached the bat's long axis by 3.7 cm. This suggests that the exertion of the adjusted internal and external torque of the right shoulder, the flexion and extension torque of the left elbow, and the left and right rotational torque of the torso, based on the dynamic contribution analysis for speed acquisition in the direction of the bat's long axis, have influenced the movements of each joint during the swing, indicating potential improvements in swing motion to respond to inside pitches.

This paper presents a novel question answering (QA) benchmark designed to evaluate contextual rule reasoning in baseball. The benchmark consists of a wide range of questions derived from play-by-play scenarios, requiring judgments on rule applications such as determining whether a run is scored or an out is made. Two large language models (LLMs) are evaluated on this benchmark to assess their accuracy. Furthermore, this study also investigates the impact of prompt structures, such as few-shot prompting, on model accuracy. Through this research, we will clarify the actual reasoning abilities of LLM in the field of baseball when faced with questions involving complex rules, and indicate future application possibilities and directions for improvement.

In soft tennis coaching, skill acquisition often relies on instructors’ sensory-based feedback. However, discrepancies in sensory perception between experienced and novice players can lead to miscommunication. For example, novices may misinterpret verbal cues and incorrectly adjust racket angles during strokes. While high-speed cameras offer objective motion analysis, their cost and complexity limit accessibility. This study aims to investigate the relationship between swing motion and grip force characteristics as a means of predicting swing trajectories. Four participants—two skilled players and two beginners—performed return shots under match-like conditions. Swing motions were recorded using a high-speed camera, and grip force data were collected using a custom-built grip sensor. The analysis revealed distinct differences in swing motion and grip force characteristics between skilled and novice players. Furthermore, a correlation between swing motion and grip behavior was identified. These findings suggest that grip sensors could provide a accessible method for quantifying swing characteristics in soft tennis, potentially enhancing both coaching and training.

As the saying goes, “the eyes are as eloquent as the mouth,” gaze behavior reflects what a person focuses on during cognitive processing. In sports science, eye-tracking data has long been valued in performance and expertise research. This study examined the gaze behavior of baseball batters using a wearable eye-tracking device, Tobii Pro Glasses 3. In a realistic batting scenario simulating actual game innings, pitches were delivered by both a pitcher and a catcher. The experiment involved three experienced players and one inexperienced participant. Heatmaps and questionnaire responses revealed that the inexperienced participant tended to focus more on the ball itself. Additionally, the questionnaire suggested that information from analysts can influence gaze behavior during batting. Experienced players also showed a tendency to become more aware of ball trajectories as they accumulated pitcher-related knowledge and gained more head-to-head experience. These findings highlight the cognitive differences in visual attention between novice and skilled players.

This paper proposes a new method for 3D velocity estimation by RTK-GNSS and inertial and magnetic field sensors. This method was used Extended Kalman filter for the posture estimation and the 3D velocity estimation. The 3D estimation was established for performance improvement in the previous proposed method, and the calibration parameter by magnetic disturbance in the magnetic field sensor output was compensated by RTK-GNSS output. The measurement experiment was conducted by a short sprint running. The GNSS and MIMU were attached to top of the head of a sprinter. The estimation results represented the 3D velocity indicating the change by break and propulsion forces, and this method represented more stable results than previous proposed method.

This study presents a method for sensor calibration under dynamic loading conditions, specifically targeting sensor clubs. In particular, forces and moments at the strain gauge attachment points under arbitrary swinging loads are calculated using inverse dynamics. Using these physical quantities, conversion equations from multiple strain signals to various physical parameters are determined [4]. Here, as an example, we present a calibration case involving kinetic variables using both hands, rather than separate calibrations for each hand.

This study proposes a new motion analysis method by combining inertial sensors, insole-type foot pressure sensors, and the method the force plates for measurement of ground reaction force were used for kinematic and kinetic analysis. This method can obtain information for calculation of joint torque. The center of pressure (COP) by measuring the foot pressure sensors of insole type was compensated using the rotational matrix. The compensated COP and the force plates output were used for this analysis. The measurement experiment was conducted by a basketball free throw motion, the analysis by the ground reaction force, foot pressure sensor output. The results indicated the COP by force plate and foot pressure sensor, and the effectiveness of this method for motion analysis with inertial sensors.

This study proposes a new analysis method for quantifying ankle joint torque during basketball shooting by combining insole-type foot pressure sensors, inertial measurement units, and two high-performance force plates. This method enables the collection of both kinetic and kinematic synchronized data from both feet, without optical motion capture system. The center of pressure (COP) data measured by the foot pressure sensors was compensated to the coordinate system of the force plates, and foot segment kinematic parameters were estimated from IMU data using Extended Kalman filter. The ankle joint torques were calculated for each foot using inverse dynamics analysis by integrating these data with ground reaction forces. In the experiment, a skilled male basketball player conducted multiple free-throw attempts on two force plates, with successful shots selected for analysis. The results represented the detailed evaluation of ankle joint torques throughout the free-throw motion, revealing characteristics between right and left feet, and trial differences. This method offers a simple and effective solution for capturing dynamic ankle joint function in practical basketball environments.

Evaluating sports performance requires effective imaging capabilities to capture high-speed movements of players and equipment over extended periods. In particular, the shuttlecock moves at very high speeds and follows an unpredictable trajectory, which poses significant challenges for effective capture. Consequently, we explored the use of an event camera. An event camera efficiently measures brightness changes with microsecond accuracy under high-speed motion while using lower memory consumption. This camera enables users to continuously monitor their performance over extended periods. This study presents a method that utilizes an event camera to estimate the speed of shuttlecocks in badminton smashes and evaluates the accuracy of this method. Uniquely, it estimates the impact time used for shuttlecock speed estimation solely based on data captured by the event camera. As a result of estimating the shuttlecock speed for one male and one female player, a significant difference was observed, demonstrating the utility of this method. As a limitation, the proposed method relies on an ideal filming environment, including the use of a black curtain for the background and a solid white frame racket.

Using 8 smart phones plus 8 external live cameras, we provide a live video streaming system. The point is our time synchronization system with the accuracy of 10msec from 150km/h moving 8 motorcycles via mobile telecommunication network. This system might be a base technology for sports video camera solutions.

Estimating the timing of impact during batting is essential for supporting skill improvement and advancing motion analysis in sports. However, conventional RGB cameras face two significant challenges: achieving high temporal resolution typically requires expensive equipment, and capturing stable measurements becomes difficult under low-light or noisy conditions. To overcome these limitations, event cameras have emerged as a promising alternative. Unlike traditional cameras, event cameras asynchronously detect brightness changes with millisecond-level precision and exhibit strong robustness in challenging lighting environments. In this study, we propose a novel method for estimating the impact timing in batting using an event camera. By leveraging the high temporal resolution and noise resistance of event-based sensing, our approach enables accurate and reliable detection of impact events. The results demonstrate the potential of event cameras as a powerful tool for sports performance analysis and open new avenues for their application in sports science.

In this paper, we propose a novel method for player identification in soccer videos. While conventional jersey number recognition cannot be applied during training sessions, our approach identifies players by first detecting walking segments using action recognition and then applying gait recognition. This method is effective even in situations where jersey numbers are not visible, such as training sessions, warm-ups, and matches without numbered uniforms. Experimental evaluation using futsal video footage demonstrated the effectiveness of the proposed approach. In the future, we aim to further improve accuracy through advancements in silhouette extraction and deep learning models.

This study proposes an automatic lap time measurement method for the Team Pursuit event on a 250 m indoor cycling track. Rider detection is performed by a fine‐tuned YOLO model, and consistent identity assignment is maintained via ByteTrack. Homography transformation using OpenCV allows precise mapping of the front‐wheel tip position to the finish line. To mitigate detection jitter, a correction algorithm assumes equal average time over each quarter lap (≈ 3.75 s). Evaluated on four overseas competition videos, the proposed system achieved average timing errors below 0.05 s, substantially reducing manual analysis effort. Future work will incorporate dynamic correction models that account for velocity variations

This paper presents an automated basketball shot detection system that uses YOLO (You Only Look Once) object detection and computer vision techniques to track the ball and hoop in real-time video footage. The system analyzes trajectory patterns to classify shot attempts and determine successful makes or misses. By employing motion filtering, geometric heuristics, and temporal analysis, the algorithm achieves robust performance in dynamic game environments. Results demonstrate accurate shot logging with real-time overlay visualization, providing a foundation for sports analytics applications.

In Japan, the use of ICT in field sports is progressing. Especially in soccer, player positions have been applied to tactical analysis. However, few research has focused on the ball, which is important for understanding the game content. Due to the small size of the ball and its occlusion with players, it is difficult to detect the ball from the video. Therefore, we propose ball detection method based on deep learning and iterative learning using video captured by two fixed 4K cameras. First, a model is built with first dataset containing balls and players. Next, we manually extract images in which balls are correctly detected from the detection results to make a second dataset. Experiment confirm that the detection accuracy improves with iterative learning, demonstrating that the proposed method is useful as a general-purpose ball detection method.

Introduction:

The shoulder joint plays a key role in sports but is prone to injuries from overuse. Muscle fatigue, particularly in the external rotators, is a common cause of dysfunction. This study examines how pitching-induced fatigue affects the cartilage contact surface of the shoulder joint using MRI.

Methods:

Four male baseball pitchers participated. Before pitching, MRI scans were performed in three shoulder positions: neutral, 120° abduction, and 120° abduction with maximum external rotation. To induce fatigue, participants performed 100 repetitions of shadow pitching. After pitching, an additional MRI scan was conducted in the maximum external rotation position. Muscle fatigue was assessed using electromyography (EMG) and measurements of maximum muscle strength for the infraspinatus, deltoid, and upper trapezius muscles.

Results:

Contact area changes after pitching varied among participants. Infraspinatus fatigue correlated with a decrease in contact area, while those without significant fatigue showed increased contact. All subjects exhibited deltoid fatigue, likely due to consistent abduction during pitching. Muscle strength decreased in some participants post-pitching, with individual differences.

Discussion:

The decrease in contact area likely reflects reduced stabilizing function of the infraspinatus, a key rotator cuff muscle. Differences in pitching form may explain variations in fatigue and contact surface patterns.

Conclusion:

Infraspinatus fatigue caused by pitching may alter joint contact mechanics, potentially leading to shoulder injuries. These findings can inform injury prevention and rehabilitation strategies for athletes.

This study presents a vision-based analysis system for estimating the sweeping position and direction of a curling brush using a monocular camera. The system detects brush regions as oriented bounding boxes and transforms image coordinates into sheet coordinates via calibration. Both positional and angular estimations are quantitatively evaluated against motion capture data. The system achieved an average position error of approximately 0.05 ± 0.03m and an angular error of around 6.55 ± 5.48°, particularly accurate at closer distances. These results suggest that the system is effective in visualizing sweeping behavior and holds promise for use in technical coaching, performance feedback, and strategic decision-making in curling.

This study focused on the observational learning process in imitation learning and investigated the effects on their action recognition when table tennis players were given video images with different characteristics. Fourteen right-handed table tennis players were filmed performing the 'counter-rotational forehand drive' action, and their body feature points were estimated using MediaPipePose. In addition to the visible light images, the original images were presented to the players as silhouettes and images of only the body feature points, for a total of three types of images, and the similarity between the questionnaire and the movements was analysed. As a result, it became clear that the viewpoint moved away from the hitting point as the amount of image information decreased.

In blind football, players compete without visual information, making movement accuracy based on their sense of proprioception crucial. During matches, players rely on auditory information to perceive their surroundings and move accordingly. While research has been conducted on the sound localization abilities of blind football players, little research has been done on their movement. This study focuses on movement accuracy. The participants were asked to identify the space they should move to based on auditory information, and then they physically moved towards that space. The movement accuracy was compared between conditions with and without a sound source, evaluating the impact of sound presence on the players' movement. The results showed that, compared to novices, experienced players had higher movement accuracy when sound was present. A comparison between the sound and no-sound conditions revealed that experienced sighted players had significantly higher movement accuracy when sound was present compared to when it was absent. On the other hand, for experienced visually impaired players, no significant effect of sound presence or absence on movement accuracy was observed. In the future, we hope to deepen the understanding of the movement mechanisms of visually impaired individuals and provide valuable insights that contribute not only to blind football but also to the field of rehabilitation.

As the real world changes from moment to moment, we live in the "now or never" world. Movement is essential to life. That is why we living things are called “creatures”. We create movement to survive. Sports are interesting because they help us enjoy the real world. We need to create movement. But as the real world is changing constantly, we need to adapt to these changes. Thus, what is needed in sports is not how, i.e., tactics, but what decisions we should make is important. Yes, strategy plays an important role in sports just in the same way we need them in our real life. However, as the real world is changing continuously, the world we come across is the one we have never experienced. The only tool we have to overcome this challenge is our instincts. In this paper, we propose an approach to enhance our instinct. Then, we can enjoy sports and our daily life more and more.

Previous research in behavioral economics has indicated that cognitive biases—systematic deviations from rational judgment—are less likely to occur in contexts involving high rewards and competitive pressure. Motivated by this framework, recent studies have examined the presence or absence of such biases in sports settings. In the case of soccer penalty shootouts, it has been observed that goalkeepers tend to dive to the left or right rather than remain in the center, despite the latter being statistically associated with higher saving rates. This pattern of behavior has been attributed to action bias, a bias toward action driven by the desire to reduce post-decision regret. The present study investigates whether action bias is also observed in eSports-based penalty kick scenarios by comparing goalkeeper behavior in penalty shootouts from the FIFA World Cup and from international eSports tournaments. This study analyzed the frequency of center-staying versus lateral-diving actions among goalkeepers in both settings using video data. The results revealed that, in the World Cup data, goalkeepers exhibited a strong tendency to dive, consistent with action bias. In contrast, this behavioral tendency was not observed in the eSports data. These findings suggest that action bias may be reduced in environments where physical exertion and social pressure are reduced, offering new insights into the situational conditions that influence decision-making biases.

To evaluate the human balance ability during standing, a system identification method involving the application of small perturbations to the subject was employed. In previous studies, we proposed a method that utilizes the horizontal sway of a support surface. However, it remains unclear whether the assumed shift in the target position of the center of mass (COM) in response to the velocity of the support surface is attributable to visual feedback. This study aimed to investigate the changes in the feedback system under different visual conditions by applying external stimuli under three conditions: eyes open, upper visual field only, and eyes closed. Eight healthy males participated in this study. Our results showed that the frequency response function varied significantly below 0.25 Hz across the three visual conditions. Notably, in the upper visual field condition, control performance of the delayed state feedback system improved, although a clear reduction in visual feedback weight was not observed. These findings suggest that restricting the visual field may prompt individuals to enhance their balance stability more on maintaining balance.

It is known that a phenomenon in which a cardiac rhythm synchronizes with a locomotor rhythm occurs during running. At this synchronization, the cardiac rhythm rate increases in proportion to the locomotor rhythm intensity, and when it approaches the limit locomotor rhythm intensity, an unstable phenomenon of the cardiac rhythm increases becomes sometimes observed. This instability of the cardiac rhythm causes lower exercise efficiency, so it is necessary to keep the stable cardiac rhythm up to the faster locomotor rhythm as far as possible. Therefore, we obtain a model with a locomotor rhythm added to the van der Pol equation that expresses the cardiac rhythm, and propose a reset control of a central command to suppress the instability of the cardiac rhythm for this model. By simulation, we show the reset control of the central command to suppress this instability.

In curling, accurately predicting stone trajectory and collision dynamics is a critical factor that directly influences decision-making and tactical strategies during games. However, conventional physical models have limitations in accurately reproducing the momentum transfer involving the rotation of the stone and the friction with the ice, resulting in significant prediction errors. In this study, we construct machine learning models that predict the trajectory and collision dynamics of stones using actual measurement data, and we propose a method that achieves higher prediction accuracy than conventional physical models. We conducted 114 collision experiments on a curling sheet using motion capture at 100 fps, attaching six markers to each stone to obtain center positions, handle angles, velocities, and angular velocities. Preprocessing included outlier correction via linear interpolation and fourth-order Butterworth low-pass filtering (0.2 Hz cutoff, 10 Hz sampling). The physical model, based on Uehara and Ito’s digital curling simulator, assumes equal masses, perfectly elastic collisions and frictional deceleration. The machine learning models comprise a Random Forest regressor for collision prediction and a Huber-loss regression for trajectory prediction, both using standardized features and horizontal-flip data augmentation. We then integrated the trajectory and collision prediction models to simulate the positions at which two stones would stop or go out based on actual measurement data from a single frame near the hog line. Comparison of the machine learning models and physical models showed that the mean absolute error (MAE) and standard deviation (SD) were reduced in most metrics for the collision models, the trajectory models, and the integrated simulations. These results suggest that our method can provide more accurate information to inform tactical decision-making and pre-match simulations in curling. Future work will focus on expanding the dataset and further refining the models toward the development of a practical curling-strategy simulator.

This paper presents a lightweight, real-time system for tracking and visualizing curling stone trajectories using consumer-grade mobile devices such as iPhones and iPads. The system consists of two components: Curl Tracer, which utilizes a YOLO-based model to detect and track stones, and Curl Viewer, which receives and visualizes the tracking data in real-time. Deploying multiple Tracers reduces occlusion-related errors and enables stable, high-precision tracking. The system supports both wireless and wired communication, allowing reliable operation in environments where wireless signals may be restricted. It provides immediate visual and quantitative feedback, making it suitable for both training and competition scenarios. The proposed system contributes to the practical deployment of data-driven tactical analysis tools in curling and other precision sports.

In this study, we aim to improve the prediction accuracy of expected score distributions in the context of digital curling. To this end, we incorporate features derived from a win probability table into a Transformer-based model. A win probability table provides the statistical win rate for each combination of point difference and the number of remaining ends, serving as contextual information that complements stone placement data. In prior research, game context was represented only by one-hot vectors of point difference and end number, without explicitly utilizing strategic statistics. However, the shape and bias of expected score distributions are likely influenced by such higher-level information. Therefore, we propose a method to combine the win probability table, encoded as a 17-dimensional vector, with the conventional game information and stone features as model input. We constructed a dataset with 95 variations of game situations for each stone configuration and generated expected score distributions via stochastic simulation of final shots. Experimental results show that the proposed model consistently achieves lower validation and test losses than the baseline. This suggests that incorporating win-rate–based features contributes to enhanced prediction accuracy and generalization performance.

This study proposes several refinements to a deep-learning-based player AI for Digital Curling 3 (DC3). Building on the baseline KuraCurling system, we first replace low-quality self-play records with 10000 games generated by the three strongest AIs from the 2024 Digital Curling Competition, enhancing both the quality and quantity of supervised data. Second, the network’s input planes are extended with end-planning cues—score differential, current end, and hammer possession—so the model can internalize long-range strategic value. Third, we train dedicated policy networks for hammer and non-hammer situations to capture their asymmetric tactics. The 100-game evaluation matches our best model, JS-model (10 000), which achieves a 0.93 win rate against the original KuraCurling. Adding end-planning features and network separation boosts the win rate by up to 10 percentage points when sufficient data is available. These results demonstrate that carefully curated teacher data, enriched state representations, and context-specific policies substantially increase DC3 performance. Future work will incorporate a value head and Monte-Carlo Tree Search to exploit search-based optimization further.