通过跟踪数据的应用, 数据执行的ai衍生的形状分析模型正在改变vip威尼斯登录入口识别团队形状的方式, 在和不在占有, 在比赛中.


作者:Jonny Whitmore, Thomas Seidl


Formations are concepts that every football fan understands and are a common topic of discussion, 无论是典型的4-4-2阵型, 控球制为主的4-2-3-1阵型或一直流行的后防线三人阵型. 但这真的能告诉vip威尼斯登录入口你的团队是如何运作的吗? 瓜迪奥拉的4-3-3阵型和Jürgen克洛普的4-3-3阵型一样吗?

Sky Sports pundit Jamie Carragher recently spoke on Sky Sports’ Monday Night 足球 about the need to look at formations 在和不在占有 and the manual process undertaken to recognise these distinctions.

问题是,地层分析仍然是一项主观和耗时的任务, with experts like Carragher having to spend their mornings going back through video to figure this out. 也就是说,直到现在……

使用Stats执行的专有形状分析, we’re able to use tracking data and the latest machine learning techniques to dive into the finer details of what a team’s shape really looks like, 无论是在押还是在押, and iron out some of the weaknesses of more traditional measures such as starting formations and average positions. vip威尼斯登录入口可以在周一早上帮杰米完成任务.


Tracking data from team sports is inherently unordered as players constantly move around the pitch. 数据执行’s 形状分析 model uncovers the strategy behind this movement by considering how players are positioned relative to each other at any given moment in the game. 通过区分一支球队是在控球状态还是不在控球状态, 这让vip威尼斯登录入口能够区分它们的进攻和防守形状.

为了将这些关系与足球中使用的已知形状相匹配, vip威尼斯登录入口使用超过2的层次聚类,当一支球队拥有17种不同的形状时,000个游戏追踪数据来识别, 当球队失去控球权时,有13种不同的形状. 使用这种数据优先的方法, we aren’t biased by the traditional formations labels and we can identify the shapes most commonly used by teams.

To detect a team’s shape (and changes in shape) within a game, we follow a three-step approach:

  • 当一支球队在控球或不在控球时,将比赛分成不同的间隔.
  • 在这些间隔, we use an unsupervised machine learning technique to identify a shape for the team 在和不在占有, 分别.
  • Assign the most likely shape from our templates (17 in possession and 13 out of possession) to the shape recognised within the interval.

不同的占有间隔自然地被比赛中的停顿所打破.g., 替换, 进球和红牌)或失误,因为这些都是教练改变战术的主要因素. Each interval must contain a minimum amount of ball in-play time in possession in order to allow for sufficient data.


Let’s look at the traditional ways of measuring team structure and then see how 形状分析 tells the story during Thomas Tuchel’s first Premier League game in charge against Wolverhampton Wanderers.


最常见的描述一个球队的形状是开始阵型或阵容图, typically seen in post-match reports or one hour prior to kick-off with the release of the team news (as avid 幻想 Premier League users know all too well). These formations have to be manually assigned or anticipated by an analyst watching the game because, 意料之中的是, 经理们很少愿意,也很少被要求发表这样的战术见解.

那么托马斯·图切尔在他的第一场比赛中是如何建立的呢? He surprised many by shifting from Frank Lampard’s favoured back four to a “Conte-era” style back three, starting with a 3-4-3 formation and utilising Ben Chilwell and Callum Hudson-Odoi as his wing backs:

Line up graphics like this are a very static representation of a team’s shape and anyone who’s watched a Pep Guardiola team during the last few years knows that these formations can be far more fluid in reality. These formations struggle to bring out the nuances of a team’s true shape during the game and can be biased by the assigning of traditional player positions. 卡勒姆·哈德森-奥多尼会扮演和本·奇尔维尔一样的边后卫角色吗?


这就是平均位置图派上用场的地方. 自动化的数据, these are commonly used by broadcasters and give us additional insight into the positions that players had within these formations. Chelsea’s average positions against Wolves suggest that Hudson-Odoi was the far more advanced wing back and that the two forwards either side of Olivier Giroud played very narrow.

与大多数人的预期相反, average position graphics are typically based on the average location of a player’s touches rather than the physical locations of the players (we’ll save ruining what ‘touches’ are for another time). 当然, 当可用的, tracking data can be used in these average position graphics to more accurately portray player positions, 但这两种方法都有其局限性:

  • 如何准确区分占有和非占有的位置?
  • 如果玩家的位置在游戏中发生变化会发生什么.g.,边锋交换位置)?
  • 如果球队在比赛中改变阵型会发生什么?
  • vip威尼斯登录入口如何在游戏后期准确地显示替补队员的位置?

而基于跟踪数据的方法可以缓解持有问题, you can still see how some of the above issues are evident when we focus on Wolves’ average positions:

从阵型图(3-4-1-2)来看,阵型基本符合预期。, 狼队三名前锋的平均位置似乎都在同一个地方, 特别是在不占有的时候. vip威尼斯登录入口知道这并不能准确反映狼队在比赛中的真实状态.


使用形状分析输出切尔西对狼, vip威尼斯登录入口现在可以清楚地识别形状, 无论是在押还是在押, 每个队在比赛中最常使用的是:

为切尔西, we can see that their wing backs push much further forward while the team is in possession and that the wide forwards remain fairly narrow.

虽然这些与vip威尼斯登录入口通过使用平均位置发现的结果相似, we can now also understand the locations of the three Wolves forwards more accurately by assigning the locations of each player within the identified shapes.

与形状分析, we can see that the average positions were misleading for Wolverhampton’s three forwards because they were rotating during the game. 虽然这似乎是显而易见的,但vip威尼斯登录入口现在可以量化它. Daniel Podence occupied the most central of the forward positions in this shape 56% of the time, but both Pedro Neto and Adama Traore also featured there during the game (16% and 13% of the time 分别). 形状分析框架也包含了替代品, 威廉José(他在72年的比赛中脱颖而出nd 分钟)在这个中心位置打了总时间的11%.

This is confirmed in the screenshots below as you can clearly see that the Wolves forwards were in their expected positions in the fifth minute but, 十五分钟后, 已经交换了. 这在比赛中很常见.

在检测团队最常用的形状的基础上, 通过将模型应用于不同的时间区间, 它还可以让vip威尼斯登录入口在游戏中检测到团队的形状变化. 这是目前只有在游戏中由分析师手动记录时才可能实现的, 但在分析中,这是一个重要的战术转变.


通过这种数据导向的方法, 形状分析 has a number of scalable applications which can save huge amounts of time for analysts in the industry and improve story-telling capabilities to engage fans:

  • 在托马斯·图切尔执教的前10场比赛中,切尔西的状态是怎样的?
  • How did Brendan Rodgers change Leicester’s shape to mastermind their victory against Liverpool?
  • 阿森纳的布卡约·萨卡在比赛中轮换位置的频率是多少?
  • 瓜迪奥拉带领的曼城在控球内外的表现如何?
  • 曼联最有效的进球形态是什么?

形状分析 automatically identifies the shape of a team (and changes of shape) within a single game that can be applied across multiple games to provide actionable insights. 构造是最初的构建模块,但形状组成了一个更具描述性的故事.