Scholarlink

<p>It has become common practice in topflight leagues to track position data of soccer players and the ball. Analyzing sports performance based on this high-resolution data is a non-trivial task due to the great complexity and simultaneous lack of structure of the game. Sports practitioners tackle this problem through <i>tactical periodization</i>, i.e., mapping the course of the game onto different states, so-called <i>match phases</i>. However, creating manual <i>tactical periodizations</i> is a time-consuming task prone to subjective biases. Automatic approaches are thus preferred, but validated and open <i>match phase</i> models are currently lacking. The present study addresses this issue by (1)&#xa0;formalizing a domain-specific, qualitative <i>match phase</i> annotation scheme from related work, (2)&#xa0;creating and validating a multi-annotator set of annotations, (3)&#xa0;training several supervised machine learning architectures to fully automate the task of annotation, and (4)&#xa0;demonstrating the usefulness by conducting a contextualized detection of playing formations with the best model, referred to as FeatGRU. Steps (2)&#xa0;through&#xa0;(4) were performed on a set of real-world soccer data and the best-performing model is made available. FeatGRU is of value to the soccer community as it provides a fully automatic, frame-by-frame <i>match phase</i> annotation that matches domain experts’ opinions with 80% accuracy while being modular extendable for future work. Moreover, we found a strong relation between semantic complexity of <i>matchphases</i>, expert agreements, and classification performance, highlighting the importance of valid label generation. Thus, our approach presents an interesting benchmark to domains where automatic approaches are required while ambiguity between human expert opinions exists.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Contextualization of soccer analysis with tactical periodization and machine learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年3月25日

查看全文
  • Henrik Biermann,
  • Daniel Memmert,
  • Niklas Petersen,
  • Dominik Raabe

摘要

It has become common practice in topflight leagues to track position data of soccer players and the ball. Analyzing sports performance based on this high-resolution data is a non-trivial task due to the great complexity and simultaneous lack of structure of the game. Sports practitioners tackle this problem through tactical periodization, i.e., mapping the course of the game onto different states, so-called match phases. However, creating manual tactical periodizations is a time-consuming task prone to subjective biases. Automatic approaches are thus preferred, but validated and open match phase models are currently lacking. The present study addresses this issue by (1) formalizing a domain-specific, qualitative match phase annotation scheme from related work, (2) creating and validating a multi-annotator set of annotations, (3) training several supervised machine learning architectures to fully automate the task of annotation, and (4) demonstrating the usefulness by conducting a contextualized detection of playing formations with the best model, referred to as FeatGRU. Steps (2) through (4) were performed on a set of real-world soccer data and the best-performing model is made available. FeatGRU is of value to the soccer community as it provides a fully automatic, frame-by-frame match phase annotation that matches domain experts’ opinions with 80% accuracy while being modular extendable for future work. Moreover, we found a strong relation between semantic complexity of matchphases, expert agreements, and classification performance, highlighting the importance of valid label generation. Thus, our approach presents an interesting benchmark to domains where automatic approaches are required while ambiguity between human expert opinions exists.