1. Abstract

1.1 Introduction

Driblab willingness to obtain physical data comes preceded by a high demand in the football club market for physical data of players of several leagues. Typically, professional football clubs have access to their own fitness data (distance run, high speed, sprints, etc...) by GPS providers hired by the own club and, in some cases, they also have access to data from other teams within their league due to a league deal with a GPS provider, that gathers and publishes such data internally for the league. Driblab aims to provide high-quality fitness data of a wide variety of leagues to complete its leading event data products already available to clients. To achieve such goal with the highest precision possible, input data such as the video of the game and game information as players lined up or team jersey characteristics, must comply a set of quality standards and as precise as possible.

1.2 Solution

This state-of-the-art system is designed to measure and analyze key physical data of football players with precision and efficiency. Leveraging advanced algorithms, our solution provides a comprehensive platform for identifying, tracking, and calculating vital player metrics from broadcast video. This product specification document serves as a comprehensive guide to our computer vision solution, outlining its key features, functionalities and technical specifications and also offers a real case scenario in which a set of fitness parameters have been extracted successfully from a game. It is intended to provide our clients, partners, and stakeholders with a detailed understanding of the capabilities and advantages offered by Driblab’s innovative technology.

2. How

The process of creating a system capable of obtaining certain fitness characteristics from the video of a game comprises a large amount of different tasks to be able to recognise players and obtain positional information of each of them. Many challenges must be faced due to the characteristics of the video from which the system obtains all the information required, so each step of the process must be robust enough to tackle all difficulties that come from the wide variety of videos analysed. Amongst the most important tasks we can list the following:

• Recognition of direction of attack: For a better recognition of players, algorithms rely on the current formation of the team, so it becomes important to know the direction of attack of both teams
• Team Identification: Players must be clustered based on their jersey colour, so that teams are identified properly, taking into account referee and goalkeeper jerseys which are typically different from the rest of the team.
• Line Recognition: To position all players on the pitch, the system need to identify key points of the pitch to be able to create a map of it an place each player on the right position. For that, lines must be found and identified properly regardless the distance or height of the camera to the pitch, the zoom of the camera in that precise moment, or the weather conditions that can make the image difficult to analyse in some scenarios.
• Scene Recognition: We must be able to differentiate between different camera situations; when the camera focuses on one player, we lose sight of players that we were tracking in previous frames, so the system must be able to recover from these situations once the camera shot comes back to a position in which the system can detect most of the players again.
• Player Detection: An object detection model takes care of recognising players on the pitch; for this task, difficulties as occlusion between players, players around the pitch or jersey colours must be addressed to obtain good results.
• Player Identification: The system uses different techniques to identify which person of the pitch previously detected corresponds to each player; from reading jersey numbers to analyse certain characteristics of the player or position on the pitch in respect to its teammates.
• Player Tracking: Players must be not only recognised but also tracked during the duration of the game; situations as camera zooms, players getting out of the scene or occlusion of players between them make this task complicated.
• Data Imputation: Given that the video contains cuts, zoom shots or transitions in addition to the fact that some players are no visible during some camera shots, the position of missing players must be inferred. A model takes over during these situations in order to impute player positions predicting how players usually move on the pitch.

We will provide a comprehensive and detailed explanation of some of the tasks previously outlined. This section is designed to clarify the most important tasks of the process, outlining the specific procedures, methodologies, and resources required for successful execution. Our aim is to ensure a thorough understanding of the tasks, enabling efficient and effective understanding.

2.1 Player Identification & Tracking

Player Identification is one of the key phases that allow us to assign each action occurred during the game to the right player of each team.