Use this link to cite:
https://hdl.handle.net/2183/47613 Desarrollo y adaptación de técnicas de calibración de cámara en partidos de fútbol
Loading...
Identifiers
Publication date
Authors
Lojo Martínez, César
Advisors
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen]: En el contexto del fútbol profesional, el análisis automatizado de imágenes y vídeos se ha convertido en una herramienta clave para optimizar el rendimiento táctico y técnico de los equipos. La creciente disponibilidad de material audiovisual, junto con los avances en visión por computador, ha abierto nuevas posibilidades para el desarrollo de soluciones orientadas al análisis espacial y posicional del juego. Este trabajo se enfoca en perfeccionar técnicas de calibración automática de cámaras empleadas en grabaciones de fútbol profesional, cuya aplicación permite generar vistas cenitales y otras representaciones útiles para el análisis táctico, como el estudio de formaciones, mapas de calor u ocupación de espacios. Como punto de partida, se busca implementar y optimizar algoritmos de detección de puntos clave en imágenes y vídeos de partidos, con el objetivo de aplicar transformaciones geométricas precisas que permitan interpretar las escenas de cámara en relación a su posición real en el campo. Una aplicación inmediata será la generación de representaciones bidimensionales del terreno de juego a partir de vistas en perspectiva, facilitando el análisis espacial de las jugadas en un sistema de coordenadas canónicas del campo. El trabajo se centra también en su aplicación en distintos entornos reales, como la ciudad deportiva del R.C. Deportivo, donde el club busca trasladar técnicas usadas en escenarios de alta calidad (como puede ser una grabación en un estadio de fútbol profesional) a condiciones menos controladas en partidos grabados con la tecnología que emplea el propio club. Entre los retos destacan las variaciones en la iluminación o las diferencias en las posiciones de cámara. Superar estas dificultades podría abrir la puerta a sistemas de análisis más avanzados en el ámbito del fútbol profesional y acercarlos al fútbol no profesional. El trabajo se desarrolló siguiendo una metodología ágil, de tipo Scrum, basada en iteraciones y reuniones periódicas con los tutores. Se implementaron técnicas de visión por computador con Python y PyTorch, utilizando un modelo HRNet para la detección de puntos clave y el algoritmo RANSAC para el cálculo de homografías. Además, se aplicaron filtros y heurísticas de postprocesado, como detección de jugadores con YOLO y ajustes dinámicos de confianza, con el fin de mejorar la precisión en distintos escenarios reales grabados por el R.C. Deportivo.
[Abstract]: In professional football, the automated analysis of images and videos has become a key tool for optimizing the tactical and technical performance of teams. The increasing availability of audiovisual material, together with advances in computer vision, has opened new possibilities for developing solutions focused on the spatial and positional analysis of the game. This work focuses on improving automatic camera calibration techniques used in professional football recordings, enabling the generation of top-down views and other representations useful for tactical analysis, such as studying formations, heat maps, or spatial occupation. As a starting point, keypoint detection algorithms are implemented and optimized in match images and videos, with the goal of applying precise geometric transformations that allow camera scenes to be interpreted in relation to their real position on the field. An immediate application is the generation of two-dimensional representations of the pitch from perspective views, facilitating spatial analysis of plays within a canonical field coordinate system. The project also explores its application in real environments, such as the R.C. Deportivo training facilities, where the club aims to transfer techniques used in high-quality (such as a recording in a professional football stadium) settings to less controlled conditions in matches recorded with its own technology. The main challenges include variations in lighting and camera positioning. Overcoming these difficulties could pave the way for more advanced analysis systems in professional football and bring them closer to non-professional football. The work was developed following an agile methodology based on the Scrum framework, with iterative cycles and regular meetings with the supervisors. Computer vision techniques were implemented using Python and PyTorch, employing the HRNet model for keypoint detection and the RANSAC algorithm for homography estimation. Additionally, post-processing filters and heuristics, such as player detection using YOLO and dynamic confidence adjustments, were applied to improve accuracy across different real-world scenarios recorded by R.C. Deportivo.
[Abstract]: In professional football, the automated analysis of images and videos has become a key tool for optimizing the tactical and technical performance of teams. The increasing availability of audiovisual material, together with advances in computer vision, has opened new possibilities for developing solutions focused on the spatial and positional analysis of the game. This work focuses on improving automatic camera calibration techniques used in professional football recordings, enabling the generation of top-down views and other representations useful for tactical analysis, such as studying formations, heat maps, or spatial occupation. As a starting point, keypoint detection algorithms are implemented and optimized in match images and videos, with the goal of applying precise geometric transformations that allow camera scenes to be interpreted in relation to their real position on the field. An immediate application is the generation of two-dimensional representations of the pitch from perspective views, facilitating spatial analysis of plays within a canonical field coordinate system. The project also explores its application in real environments, such as the R.C. Deportivo training facilities, where the club aims to transfer techniques used in high-quality (such as a recording in a professional football stadium) settings to less controlled conditions in matches recorded with its own technology. The main challenges include variations in lighting and camera positioning. Overcoming these difficulties could pave the way for more advanced analysis systems in professional football and bring them closer to non-professional football. The work was developed following an agile methodology based on the Scrum framework, with iterative cycles and regular meetings with the supervisors. Computer vision techniques were implemented using Python and PyTorch, employing the HRNet model for keypoint detection and the RANSAC algorithm for homography estimation. Additionally, post-processing filters and heuristics, such as player detection using YOLO and dynamic confidence adjustments, were applied to improve accuracy across different real-world scenarios recorded by R.C. Deportivo.
Description
Editor version
Rights
Attribution-NonCommercial-NoDerivatives 4.0 International








