Yannis Maag

Advisors

Rebecca Lennartz (M. Sc.),  Maike Stöve (M. Sc.), Burkhard Duemler (adidas AG), Prof. Dr. Björn Eskofier

Duration

05 / 2024 – 10 / 2024

Abstract

Performance analysis has become increasingly important for the evaluation of players and games in today’s sport. In soccer as well as in sports such as tennis, basketball, and volleyball, performance analyses are already an essential component. In professional soccer, tools like chest straps and video analysis are used to monitor parameters such as the player’s position, distance covered, speed, power, intensity, and heart rate in order to optimize performance, support player health, and refine tactical approaches for future games [1–3].

Given the considerable financial resources available in professional soccer, professional clubs can afford to invest in this expensive equipment [1]. However, for amateur athletes with limited budgets, the question arises: How can they benefit from similar insights without the need for big investments into expensive technology?

A common method is to use data from an Inertial Measurement Unit (IMU), which is an ideal solution to this problem due to its small size, low cost, and high accuracy [3]. This makes data acquisition in Human Activity Recognition (HAR) highly convenient, as the sensor has minimal impact on the user [4]. It has been shown that machine learning algorithms can use IMU data recorded during a soccer drill, placed on the players’ foot or hand, allowing them to distinguish between general activities such as jogging, sprinting, passing, shooting, and jumping [1, 3, 5, 6]. Traditional machine learning methods, as used by Hossain et al., achieved an accuracy of 86.5% in classifying these activities. In contrast, Cuperman et al. reached an accuracy of 98.3% using a deep learning (DL) approach, highlighting the potential of DL in activity recognition.

Despite these advancements, existing methods lack the detailed metrics necessary for individual assessment, such as identifying the type of pass or determining if a player is dribbling while jogging [3]. Additionally, current approaches predominantly use shallow classification structures. This makes the development process heavily manual, subjective, and extremely time-consuming [3]. Although there are some deep learning approaches that aim to counteract this problem, they tend to cover general activities and have so far used little real game data in their training. Furthermore, deep learning has not been used frequently, as the limited amount of data restricts its applicability [5, 7, 8].

Therefore, the aim of this work is to develop a deep learning approach capable of classifying more specific metrics such as dribbling, type of pass, and automatic left and right foot recognition using two IMU sensors embedded in the sole of each soccer shoe. The underlying research questions are i) can deep learning algorithms classify activity metrics in soccer? and ii) can the analyzed algorithms continue to recognize the metrics even in game-like situations?

References

[1] Yuki Kondo, Shun Ishii, Hikari Aoyagi, Tahera Hossain, Anna Yokokubo, and Guillaume Lopez. FootbSense: Soccer Moves Identification Using a. Single IMU. In Md Atiqur Rahman Ahad, Sozo Inoue, Daniel Roggen, and Kaori Fujinami, editors, Sensor- and Video-Based Activity and Behavior Computing, pages 115–131, Singapore, 2022. Springer Nature.

[2] Alexander Hoelzemann, Julia Lee Romero, Marius Bock, Kristof Van Laerhoven, and Qin Lv. Hang-time har: A benchmark dataset for basketball activity recognition using wrist-worn inertial sensors. Sensors, 23(13), 2023.

[3] Rafael Cuperman, Kaspar M. B. Jansen, and Micha.‚ G. Ciszewski. An End-to-End Deep Learning Pipeline for Football Activity Recognition Based on Wearable Acceleration Sensors. Sensors (Basel, Switzerland), 22(4):1347, February 2022.

[4] Umakorn Manupibul, Rataya Tanthuwapathom, Wijit Jarumethitanont, Phawin Kaimuk, Weerawat Limroongreungrat, and Weerachai Charoensuk. Integration of force and imu sensors for developing low-cost portable gait measurement system in lower extremities. Sci Rep, 13(1):10653, Jun 2023.

[5] Maike Stoeve, Dominik Schuldhaus, Axel Gamp, Constantin Zwick, and Bjoern M. Eskofier. From the laboratory to the field: Imu-based shot and pass detection in football training and game scenarios using deep learning. Sensors, 21(9), 2021.

[6] H M Sajjad Hossain, Md Abdullah Al Hafiz Khan, and Nirmalya Roy. Soccermate: A personal soccer attribute profiler using wearables. pages 164–169, 03 2017.

[7] Emily E Cust, Alice J Sweeting, Kevin Ball, and Sam Robertson. Machine and deep learning for sport-specific movement recognition: a systematic review of model development and performance. Journal of Sports Sciences, 37(5):568–600, March 2019.

[8] Dominik Schuldhaus. Human activity recognition in daily life and sports using inertial sensors. FAU Studien aus der Informatik, 8, 2019.