Thesis ALGORITMO DE VISIÓN POR COMPUTADOR PARA DETECCIÓN Y SEGUIMIENTO MÚLTIPLE DE OBJETOS APLICADO A JUGADORES DE FÚTBOL
Loading...
Date
2016-03
Journal Title
Journal ISSN
Volume Title
Program
DEPARTAMENTO DE ELECTRÓNICA. INGENIERÍA CIVIL ELECTRÓNICA – MENCIÓN COMPUTADORES
Campus
Casa Central Valparaíso
Abstract
Esta Memoria de Título se enmarca dentro del proyecto FONDECYT: “Methodology
and Applications for Incremental Behaviour Learning in Video guided by Information
Reliability”, el cual en una de sus aristas tiene el objetivo de analizar el comportamiento
de equipos deportivos para obtener análisis automatizado del rendimiento individual
y colectivo de los atletas. Dado el contexto, el objetivo de este trabajo es realizar por
medio de técnicas de Visión por Computador, la detección y seguimiento automático
de jugadores participantes de un partido de fútbol, utilizando secuencias de imágenes
obtenidas desde una cámara de video. Este algoritmo debe ser diseñado e implementado
en el “Framework VAT: Video Analysis Tool”, el cual permite desarrollar el sistema de
seguimiento de forma ordenada y modular.
Los eventos dinámicos rápidos, situaciones de ocultación complejas, gran variabilidad
en las condiciones lumínicas; las condiciones de “zoom” y posición relativa de la
cámara con respecto al campo de juego, son algunas de las restricciones que deben ser
evaluados para definir una estrategia de solución.
En primer lugar, se debe implementar una fase de Segmentación a fin de definir las
regiones de la imagen que corresponden a los objetos de interés. Luego, en la etapa
de Detección y Clasificación se utilizan modelos de objeto para buscar candidatos a
jugador. En la última etapa, que corresponde a la de Seguimiento, se busca asociar los
objetos detectados previamente, con los objetos localizados en la imagen actual.
A través de una estructura jerárquica de modelos y un algoritmo de seguimiento de
múltiples hipótesis; se logra detectar, clasificar y seguir a los jugadores presentes en
un escenario real de competencia. Se soluciona en forma básica, el problema de la
ocultación dinámica para situaciones particulares. Se tiene un algoritmo robusto para
trabajos futuros relacionados con la obtención de información estadística de alto nivel.
This Working Title is part of FONDECYT project: “Methodology and Applications for Incremental Behaviour Learning in Video guided by Information Reliability”, which in one of its topic have the aim of analyze the sport teams behavior for automated analysis of individual and collective performance of athletes that comprise. Given this context, the objetive of this work is perform by means of computer vision techniques, the detection and automatic players tracking participants of football match, using images sequences from static video camera. This algorithm must be designed and implemented in the VAT Framework: “Video Analysis Tool”, which allows developing the tracking system in an orderly and modular. Fast dynamics events, occlusion complex situations, large variability in lighting condi- tions; “zoom” aspects and relative position of the camera with respect to the field, are some of the restrictions that must be evaluated to define a proper solution strategy. First, should be setting a segmentation phase for defining interest regions in current frame which corresponding to possible players or objects. Later, in detection and classification step are selected appropiate object models for finding player candidates in the current frame. In the last step, which correspond to the tracking phase, we associate detected objects with current observations. Through a structure containing multiple object models that relate under a hierarchical tree configuration and by a multiple hypotheses tracking algorithm, it is achieved detect, classify and track players with minimum efforts reconfiguration. It is solved in basic form, the problem of dynamic occlusions for particular situations. It has a robust algorithm for future work related to collect high-level statistical information.
This Working Title is part of FONDECYT project: “Methodology and Applications for Incremental Behaviour Learning in Video guided by Information Reliability”, which in one of its topic have the aim of analyze the sport teams behavior for automated analysis of individual and collective performance of athletes that comprise. Given this context, the objetive of this work is perform by means of computer vision techniques, the detection and automatic players tracking participants of football match, using images sequences from static video camera. This algorithm must be designed and implemented in the VAT Framework: “Video Analysis Tool”, which allows developing the tracking system in an orderly and modular. Fast dynamics events, occlusion complex situations, large variability in lighting condi- tions; “zoom” aspects and relative position of the camera with respect to the field, are some of the restrictions that must be evaluated to define a proper solution strategy. First, should be setting a segmentation phase for defining interest regions in current frame which corresponding to possible players or objects. Later, in detection and classification step are selected appropiate object models for finding player candidates in the current frame. In the last step, which correspond to the tracking phase, we associate detected objects with current observations. Through a structure containing multiple object models that relate under a hierarchical tree configuration and by a multiple hypotheses tracking algorithm, it is achieved detect, classify and track players with minimum efforts reconfiguration. It is solved in basic form, the problem of dynamic occlusions for particular situations. It has a robust algorithm for future work related to collect high-level statistical information.
Description
Keywords
ANÁLISIS AUTOMATIZADO DE EVENTOS DEPORTIVOS, VISIÓN POR COMPUTADOR, SEGUIMIENTO DE OBJETOS