Subject: Balls are an essential part of play in many sports and games. In these ball sports, the position of the ball is a valuable input for tactical considerations, studies in training methodology or other analyses and evaluations that aim for a better understanding of events and actions occurring in that sports. A common tool for the determination of the balls position is the detection of balls in digital image data and videos. Even though there are approaches for the automatized detection of balls in images there is still no total confidence in these computer-based detection solutions. The human visual system is still considered as the gold standard for the ball detection in images. In most cases, the identification and accurate position determination is therefore a manual task, carried out by human operators, scientist and analysts. Problem: The manual evaluation of image data is a time consuming and therefore a resource intensive task. For studies and statistical evaluations there is a high demand for manpower resources for the manual detection of balls in images. For real time or timesensitive analysis, the manual ball detection is too slow. Image-based recognition algorithms and systems provide promising approaches and strategies for an automized detection of balls in images. Up to now, they do not represent a real alternative to manual detection. Most recognition solutions require an extensive and complex adaption to the prevailing conditions and are not able to handle changes in the general conditions and key determinants. Furthermore, in most of the cases, the detection quality is insufficient for detailed analyses. Motivation: During the last years, machine learning evolved as a source of new approaches for different recognition tasks, including also image recognition. The principle of learning from data instead of programming by hand opened doors to new and enriching solutions for automized recognition tasks. Especially neural networks achieved human-like performance in image-based recognition tasks such as image classification or pattern recognition in images. One of the most efficient and powerful types of neural networks are convolutional neural networks. The structure of these networks is substantially inspired by the organization of the animal visual cortex, which qualifies them especially for their application in image-based recognition tasks. The application of convolutional neural networks might be a powerful tool to improve the quality, efficiency and precision of ball detection in image data. Objective: This thesis introduces convolutional neural networks as a tool for the detection of balls in image data. The ball detection is exemplarily developed for the analysis of motion sequences of balls in field or indoor hockey. The objective of this thesis is to develop a conceptual framework for the implementation of a detection solution for balls, based on the application of convolutional neural networks. Main ambition is to create the theoretical and technical basis for a robust, powerful ball detection tool that works for different areas of application and varying environmental conditions. An important demand for the development of the deep learning approaches using convolutional neural networks is to reduce the required amount of manually labeled training data to a minimum. Thesis Outline: The main part of this thesis is a theoretical examination of the possible application of deep learning algorithms using convolutional neural networks for the detection of balls in images. Chapter 2 discusses the details and characteristics of the problem of ball detection for the given reference data. Appearing problems with the detection of balls are identified and requirements for a desired solution are derived. Chapter 3 briefly introduces existing approaches for the detection of balls in images. Furthermore, promising scientific papers are mentioned that applied deep learning approaches using neural networks for related tasks in image recognition. Chapter 4 contains the fundamental theoretical base for the work with artificial neural networks. Inspired by mammalian neurons, artificial neurons are introduced and their structure and characteristics are discussed. Based on single artificial neurons, artificial neural networks are introduced and discussed. Feed-forward neural networks, as the most basic and common type of neural networks, are used to explain the fundamental concepts and features of artificial neural networks. After a general discussion of learning algorithms and machine learning models, the principles of learning for neural networks are introduced. Key part of this chapter is the theoretical introduction to convolutional neural networks. In chapter 5, the introduced theoretical principles are used to develop and to evaluate different types of deep learning models for the detection of balls in images. The evaluation of different approaches leads to the construction of a final multi-column convolutional neural network architecture that can be used for ball detection in images. Chapter 6 contains the results of the application of the developed network architecture for the detection of balls on the available image data. The results are discussed in chapter 7. This thesis closes with a conclusion in chapter 8 and a outlook on possible future work building on the results of this thesis in chapter 9.
© Copyright 2018 Veröffentlicht von Universität Leipzig, Fakultät für Mathematik und Informatik, Institut für Informatik. Alle Rechte vorbehalten.