Faculty Profiles - FUJII Keisuke

Space evaluation based on pitch control using drone video in Ultimate Open Access

IWASHITA Shunsuke, SCOTT Atom, UMEMOTO Rikuhei, NING Ding, FUJII Keisuke

Proceedings of the Annual Conference of JSAI Vol. JSAI2024 ( 0 ) page： 3Xin221 - 3Xin221 2024

Language：Japanese Publishing type：Research paper (scientific journal) Publisher：The Japanese Society for Artificial Intelligence

Ultimate is a sport where two teams of seven players compete to score points by passing a disc and catching it in the end zone. Unlike other sports, Ultimate has the unique characteristic that players holding the disc cannot move, emphasizing the importance of creating space to receive passes. While there is ample data and research on space in sports like soccer and basketball, such information is lacking for Ultimate. This study aims to focus on the widely adopted 3-vs-3 format in training by many teams and evaluate the space during offensive plays. The data used in this study was obtained through filming with a drone camera, and position data was acquired through angle correction and coordinate transformations. The model employed in this study is based on a soccer's pitch control model, which indicates the probability of controlling the ball at various points on the pitch. This model is adapted to suit the rules of Ultimate, where players cannot move while holding the disc. By combining weights that take into account position and distance with the values of the pitch control, space evaluation metrics are calculated, and the relationship between scoring and the space evaluation metric immediately before a pass was analyzed.

DOI： 10.11517/pjsai.jsai2024.0_3xin221

Front line defensive evaluation based on event detection from full pitch video in soccer Open Access

IDE Kenjiro, UCHIDA Ikuma, UMEMOTO Rikuhei, FUJII Keisuke

Proceedings of the Annual Conference of JSAI Vol. JSAI2024 ( 0 ) page： 3Xin247 - 3Xin247 2024

Language：Japanese Publishing type：Research paper (scientific journal) Publisher：The Japanese Society for Artificial Intelligence

Most data analysis in soccer requires tracking data of players and ball and event data, and currently well-funded professional leagues are collecting a lot of data manually through companies. Previous research has used machine learning models to perform specific event detection from professional broadcast video, but amateur video is difficult to be analyzed due to the existence of various recording conditions. The purpose of this study is to perform specific event detection from amateur full-pitch video and to examine the validity of front-line defensive evaluation based on the detection results. In this study, two methods were used to detect events from video: the first is a deep learning-based e2e-spot method that uses video frames as input and predictive labels as output; the second is a rule-based method that uses ball velocity from ball tracking data and detects the start of a pass based on changes in velocity. We performed the fine-tuning of e2e-spot on amateur full-pitch video, was able to detect events more accurately than the rule-based method. We examined how accurately the estimated events can be used to evaluate front line defenses in soccer.

DOI： 10.11517/pjsai.jsai2024.0_3xin247

Towards the Construction of a Soccer Foundation Model Using Simulation Data Open Access

SOMEYA Taiga, SCOTT Atom James, FUJII Keisuke, AKIYAMA Hidehisa, NAKASHIMA Tomoharu, YANAKA Hitomi

Proceedings of the Annual Conference of JSAI Vol. JSAI2024 ( 0 ) page： 3I5OS27b02 - 3I5OS27b02 2024

Language：Japanese Publishing type：Research paper (scientific journal) Publisher：The Japanese Society for Artificial Intelligence

Recent advancements in large foundation models have spanned various domains, including natural language processing, autonomous driving, and multivariate time series forecasting. Meanwhile, applied sports analysis has become widespread, with a particular focus on constructing quantitative evaluation methods for players and teams through the modeling of match situations. Despite this, a soccer foundation model capable of performing various tasks within a single model remains unexplored. This study explores a potential soccer foundation model by applying a multivariate time series prediction architecture for forecasting soccer trajectory data. We propose using log data from soccer simulation leagues for training, taking into account 1) the small scale of real trajectory data and 2) the effectiveness of synthetic data in constructing foundation models as indicated by previous research. Furthermore, we evaluate the effectiveness of the embedding representations by qualitatively comparing their similarities with actual soccer trajectories, confirming their applicability in downstream tasks.

DOI： 10.11517/pjsai.jsai2024.0_3i5os27b02

Analysis of the impact of player combinations on scoring efficiency in basketball Open Access

YAMADA Kazuhiro, FUJII Keisuke

Proceedings of the Annual Conference of JSAI Vol. JSAI2024 ( 0 ) page： 1D4GS1004 - 1D4GS1004 2024

Language：Japanese Publishing type：Research paper (scientific journal) Publisher：The Japanese Society for Artificial Intelligence

In a basketball game, the players compete against each other in a five-on-five match. In particular, it is important for players with different playstyles to cooperate and score efficiently during possessions, which take place many times in one game. In a previous study, the compatibility of players was examined using clustering results based on each player's statistics, called stats, but the findings obtained were considered to be limited due to the method of selecting features that included both offense and defense. This study focuses only on offense and aims to examine more specifically the impact of player combinations on scoring efficiency. In this study, two different methods are used to capture the playstyles of players on offense: one is a newly proposed method that clusters the tendency of shots based on the Wasserstein distance, the distance between distributions, which considers the set of shots of each player as a probability distribution using shooting features created from tracking data. The other is a method for clustering players' roles in the offense, which is a modification of the existing method. By creating and interpreting a machine learning model that predicts stats representing scoring efficiency from information on lineups based on these two clusterings, new insights into the compatibility of players were obtained.

DOI： 10.11517/pjsai.jsai2024.0_1d4gs1004

Adaptive action utilization in reinforcement learning from real-world multi-agent demonstrations Open Access

FUJII Keisuke, TSUTUSI Kazushi, SCOTT Atom, NAKAHARA Hiroshi, TAKEISHI Naoya, KAWAHARA Yoshinobu

Proceedings of the Annual Conference of JSAI Vol. JSAI2024 ( 0 ) page： 1E5GS504 - 1E5GS504 2024

Language：Japanese Publishing type：Research paper (scientific journal) Publisher：The Japanese Society for Artificial Intelligence

When modeling real-world biological multi-agents with reinforcement learning, there is a domain gap between the source real-world data and the target reinforcement learning environment. Therefore, the target dynamics are adapted to the unknown source dynamics. In this study, we propose a reinforcement learning method that uses information obtained by adapting source action to target action in a supervised manner as a method for domain adaptation in multi-agent reinforcement learning from real-world demonstrations. In limited situations such as 2vs1 chase-escape, 2vs2 and 4vs8 in soccer, we show that the agent learned to imitate the demonstrations and obtain rewards compared to the baseline.

DOI： 10.11517/pjsai.jsai2024.0_1e5gs504

Quantitative Action Evaluation Metrics in Football Based on Action Selection and State Transition Probability Estimation through Language Modeling Open Access

SOMEYA Taiga, KAWAGUCHI Kohei, FUJII Keisuke

Proceedings of the Annual Conference of JSAI Vol. JSAI2024 ( 0 ) page： 2F4GS505 - 2F4GS505 2024

Language：Japanese Publishing type：Research paper (scientific journal) Publisher：The Japanese Society for Artificial Intelligence

In applied sports analytics, especially in goal-oriented sports such as soccer, the necessity to process continuous data involving multiple agents poses a significant challenge for comprehensive analysis. Existing metrics, which often rely on static and simple inputs such as the ball's coordinates and nearby statistical data, fail to account for the broader match context. Recently, there has been growing interest in Large Language Models (LLMs), recognized for their potential not only in natural language processing but also in reinforcement learning and multi-agent trajectory forecasting. However, their application in soccer analytics is still in its infancy. This study introduces an innovative approach that employs language models to predict sequences tokenizing both the observable states and selected actions of players on the field. This method aims to model the probabilities associated with players' decision-making and state transitions. Furthermore, this research pioneers a framework to calculate and evaluate the state-action value (Q-value) of each player and team, leveraging the model's outputs within a reinforcement learning context. This marks the first attempt to establish such a quantitative evaluation framework in this domain.

DOI： 10.11517/pjsai.jsai2024.0_2f4gs505

Information completion using a deep generative model for counterfactual evaluation of spatio-temporal event data in group sports Open Access

UMEMOTO Rikuhei, FUJII Keisuke

Proceedings of the Annual Conference of JSAI Vol. JSAI2024 ( 0 ) page： 3I5OS27b03 - 3I5OS27b03 2024

Language：Japanese Publishing type：Research paper (scientific journal) Publisher：The Japanese Society for Artificial Intelligence

With the development of measurement technology, people in various fields analyze multi-agent spatio-temporal data. However, since the analysis is limited to what occurred and there is a lack of information in the data, there are challenges in utilizing the results of data analysis for future decision-making and in finding solutions to data uncertainty. Hence, counterfactual and information completion play an essential role, but only some examples of applying these methods to the data exist. This paper proposes methods to evaluate behavior by considering counterfactuals and to complement information for the data. In particular, we do this for open soccer data that includes information about a more significant number of agents. The former proposes a method to evaluate team defenses by considering counterfactuals with a mathematical model that is easy for people to understand. The latter proposes a deep generative model that can complement information about players' velocities lacking in such data while considering features related to agents' interactions. We expect this research will enable more people to analyze and support players and coaches in decision-making.

DOI： 10.11517/pjsai.jsai2024.0_3i5os27b03

Dog's 3D Skeleton Reconstruction using a Moving Trainer for Analysis of Guide Dog Training. Reviewed

Ansheng Wang, Rongjin Huang, Keisuke Fujii, Shinji Tanaka, Yoshiro Matsunami, Yasutoshi Makino, Hiroyoki Shinoda

SMC page： 4594 - 4599 2024

Extracting proficiency differences and individual characteristics in golfers' swing using single-video markerless motion analysis Reviewed Open Access

Kota Yamamoto, Yumiko Hasegawa, Tomohiro Suzuki, Hiroo Suzuki, Hiroko Tanabe, Keisuke Fujii

Frontiers in Sports and Active Living Vol. 5 2023.11

Authorship：Last author Language：English Publishing type：Research paper (scientific journal) Publisher：Frontiers Media SA

In this study, we analyzed golfers' swing movement to extract differences in proficiency and individual characteristics using two-dimensional video data from a single camera. We conducted an experiment with 27 golfers who had a wide range of skill levels, using a 7-iron; we acquired video data with a camera on the sagittal plane. For data extraction, we used pose estimation (using HRNet) and object detection (using DeepLabCut) methods to extract human-joint and club-head data. We examined the relationship between proficiency and individual characteristics vis-à-vis forward tilt angle and club trajectory. The results showed that the stability and reproducibility of the forward tilt angle are characteristics of proficiency. Highly skilled golfers showed low variability and high reproducibility between trials in forward tilt angle. However, we found that club trajectory may not be a characteristic of proficiency but rather an individual characteristic. Club trajectory was divided roughly into clockwise rotation and counterclockwise rotation. Thus, the analysis based on video data from a single markerless camera enabled the extraction of the differences in proficiency and individual characteristics of golf swing. This suggests the usefulness of our system for simply evaluating golf swings and applying it to motor learning and coaching situations.

DOI： 10.3389/fspor.2023.1272038

Can golfers choose low-risk routes in steep putting based on visual feedback of ball trajectory? Reviewed International journal Open Access

Yumiko Hasegawa, Ayako Okada, Keisuke Fujii

Frontiers in Sports and Active Living Vol. 5 page： 1131390 - 1131390 2023.8

Authorship：Last author Language：English Publishing type：Research paper (scientific journal) Publisher：Frontiers Media SA

This study aims to clarify why the aiming method in golf putting in risky situations differs based on skill level. This study set up a difficult challenge (steep slopes and fast ball rolling greens), which required even professional golfers to change their aim. A total of 12 tour professionals and 12 intermediate amateurs were asked to perform a steep-slope task with no visual feedback of outcomes (no FB) followed by a task with visual feedback (with FB). The aim of the task was for the ball to enter the hole in one shot. Additionally, the participants were told that if the ball did not enter the hole, it was to at least stop as close to it as possible. The participant's aim (as an angle) and the kinematics of the putter head and ball were measured. The results indicated that professionals' highest ball trajectory points were significantly higher than that of amateurs, especially with FB. Additionally, professionals had higher ball-launch angles (the direction of the ball when the line connecting the ball and the center of the hole is 0 degrees) and lower peak putter head velocities than amateurs. Furthermore, the aim angle, indicating the golfer's decision-making, was higher for professionals under both conditions. However, even with FB, the amateurs' aim angles were lower and the difference between trials was smaller than that of professionals. Therefore, this study confirmed that the professionals made more drastic changes to their aim to find low-risk routes than the amateurs and that the amateurs’ ability to adjust their aim was lower than that of professionals. The results suggest that the reason for the amateurs' inability to find low-risk routes lies in their decision-making. The professionals found better routes; however, there were individual differences in their routes.

DOI： 10.3389/fspor.2023.1131390

PubMed

Synergizing Deep Reinforcement Learning and Biological Pursuit Behavioral Rule for Robust and Interpretable Navigation Reviewed

Kazushi Tsutsui, Kazuya Takeda, Keisuke Fujii

1st Workshop on the Synergy of Scientific and Machine Learning Modeling (SynS and ML) co-located with the International Conference on Machine Learning (ICML'23) 2023.8

Pitching strategy evaluation via stratified analysis using propensity score Reviewed Open Access

Hiroshi Nakahara, Kazuya Takeda, Keisuke Fujii

Journal of Quantitative Analysis in Sports Vol. 19 ( 2 ) page： 91 - 102 2023.5

Authorship：Last author,　Corresponding author Language：English Publishing type：Research paper (scientific journal) Publisher：Walter de Gruyter GmbH

Abstract

Recent measurement technologies enable us to analyze baseball at higher levels of complexity. There are, however, still many unclear points around pitching strategy. There are two elements that make it difficult to measure the effect of a pitching strategy. First, most public datasets do not include location data where the catcher demands a ball, which is essential information to obtain the battery’s intent. Second, there are many confounders associated with pitching/batting results when evaluating pitching strategy. We here clarify the effect of pitching attempts to a specific location, e.g., inside or outside. We employ a causal inference framework called stratified analysis using a propensity score to evaluate the effects while removing the effect of confounding factors. We use a pitch-by-pitch dataset of Japanese professional baseball games held in 2014–2019, which includes location data where the catcher demands a ball. The results reveal that an outside pitching attempt is more effective than an inside one to minimize allowed run average. In addition, the stratified analysis shows that the outside pitching attempt is effective regardless of the magnitude of the estimated batter’s ability, and the proportion of pitched inside for pitcher/batter. Our analysis provides practical insights into selecting a pitching strategy to minimize allowed runs.

DOI： 10.1515/jqas-2021-0060

Other Link： https://www.degruyter.com/document/doi/10.1515/jqas-2021-0060/pdf

Prediction of Bat Flight Path During Obstacle Avoidance by Imitation Learning Reviewed

Shoko Genda, Yu Teshima, Taku Kawamura, Daichi Ohara, Keisuke Fujii, Shizuko Hiryu

2nd International Workshop on Behavior analysis and Recognition for knowledge Discovery (BiRD 2023) in conjunction with the IEEE International Conference on Pervasive Computing and Communications (PerCom 2023) 2023.3

Biomechanical strategies to maximize gait attractiveness among women Reviewed Open Access

Hiroko Tanabe, Keisuke Fujii, Naotsugu Kaneko, Hikaru Yokoyama, Kimitaka Nakazawa

Frontiers in Sports Active Living Vol. 5 ( 1091470 ) 2023.2

Estimating the Effect of Team Hitting Strategies Using Counterfactual Virtual Simulation in Baseball, Reviewed

Hiroshi Nakahara, Kazuya Takeda, Keisuke Fujii,

International Journal of Computer Science in Sport Vol. 22 ( 1 ) page： 1 - 12 2023.1

A framework of interpretable match results prediction in football with FIFA ratings and team formation. Reviewed International journal Open Access

Calvin C K Yeung, Rory Bunker, Keisuke Fujii

PloS one Vol. 18 ( 4 ) page： e0284318 2023

Authorship：Last author,　Corresponding author Language：English Publishing type：Research paper (scientific journal)

While forecasting football match results has long been a popular topic, a practical model for football participants, such as coaches and players, has not been considered in great detail. In this study, we propose a generalized and interpretable machine learning model framework that only requires coaches' decisions and player quality features for forecasting. By further allowing the model to embed historical match statistics, features that consist of significant information, during the training process the model was practical and achieved both high performance and interpretability. Using five years of data (over 1,700 matches) from the English Premier League, our results show that our model was able to achieve high performance with an F1-score of 0.47, compared to the baseline betting odds prediction, which had an F1-score of 0.39. Moreover, our framework allows football teams to adapt for tactical decision-making, strength and weakness identification, formation and player selection, and transfer target validation. The framework in this study would have proven the feasibility of building a practical match result forecast framework and may serve to inspire future studies.

DOI： 10.1371/journal.pone.0284318

PubMed