In top-level sport, where all participants have exceptional physical and technical skills, as well as deep theoretical knowledge of their field, the gap between the best results is minimal. The winner is determined by small details that may seem insignificant to the uneducated eye, but are actually fundamental to gaining ground on others. In swimming in particular, many finals of important competitions end up with a difference of less than a tenth of a second between the leaders. The details bringing victory can be very varied because they concern the individual physique of the swimmers, their mental and physical preparation, their understanding of the swimming style of their competitors, and many other things. Understanding them is crucial to winning: this is the role of the sports coaches. They will study with finesse what can allow their swimmer to be the most efficient.The first step in the analysis of training and races is information extraction. In this thesis, we are particularly interested in swimming competitions. Our goal is to generate an automatic race report. This would free up an invaluable amount of time for coaches, and would also allow for extensive analysis of competitions. Such technology would also improve the detection of potential talent through the systematic analysis of all amateur competitions.We will focus here on video analysis, as sensors and other intrusive acquisition systems cannot be used during championships. This imposes important constraints related to the recording conditions of the videos: our methods must be robust and general. Computer vision methods will be explored to get the best out of the videos. We will also explore image analysis in a less data-dependent way than usual. Indeed, this field has progressed enormously over the last decade thanks to the development of deep learning, but it depends a lot on the quality and quantity of data. Our general problem will therefore concern the extraction of information from swimming race videos using small amounts of data. This task will be divided into three parts, each one studying a specific type of information. All results, models, and resulting databases have been published online, accessible to all.We will start by focusing on the detection of swimmers in images. This task is the most obvious to start with, because to study a swimmer on a race, we must be able to locate him. This chapter will therefore introduce a detection method specifically adapted to swimmers, as well as a dataset related to the task.Detecting swimmers on an image is a first step, but it does not give positional information in the pool. For that, we need to register the image, that is to map each point of the image to a zone of the pool. A particularly fast and very efficient method will be explained to answer this task. Another dataset will be presented.The third part of this thesis will concern the measurement of swimming cycles. The repetition of the movement being omnipresent during a race, its study is one of the most useful to perceive the quality of swimming. It is an excellent basis for measuring a swimmer's fatigue, efficiency, or technique. A general method to count cycles on a video will be presented. Specifically for swimming, we will also describe a way to locate the ends of cycles, in order to measure their individual duration with precision.All this will lead to the fusion of these methods into a single tool capable of analyzing swimming races with little or no human intervention. The different models will be arranged in such a way as to get the most out of each and to reduce their individual errors. The qualities and limits of this tool will be presented, as well as the way to measure its accuracy. Our goal is to make it efficient enough for the FFN to use it during its competitions and training in order to improve performances.
© Copyright 2022 Veröffentlicht von Ecole Centrale de Lyon. Alle Rechte vorbehalten.