diplomsko delo
Jernej Mlakar (Author), Janez Perš (Mentor), Marija Ivanovska (Co-mentor)

Abstract

Umetna inteligenca ter računalniški vid sta v športu zelo koristna in kakovostna pripomočka, saj nam skrajšata čas iskanja, zbiranja ter analiziranja podatkov. Ker pa umetna inteligenca še ni dovolj napredna za nekoliko kompleksnejše programe ter ideje, je človeško znanje trenutno še v veliki prednosti. Če si predstavljamo program, ki bi potreboval samo posnetek rokometne tekme in v hipu vrnil segmentirane podatke vseh posameznikov ter njihovih strelov, se mogoče zasnova za tak program ne zdi tako kompleksna. Ob podrobni poglobitvi v zasnovo se zavemo, da je za nastanek takšne aplikacije potrebno mnogo različnih pristopov ter vloženega dela. Ena izmed največjih osnov za takšno aplikacijo bi bila lokalizacija igralcev z globoko naučenimi algoritmi. Takšna lokalizacija je precej odvisna od človeške pomoči, saj je potrebno algoritme zaganjati fizično. Vendar je nekje potrebno začeti, saj lahko samo tako pridemo do želenih ciljev. S pomočjo različnih vnaprej pripravljenih algoritmov ter lastnega znanja lahko iz amaterskih posnetkov rokometne tekme pridobimo lokacijo igralcev na tlorisu rokometnega igrišča oziroma iz tridimenzionalnega v dvodimenzionalni prostor. Rezultati, ki jih vrne lokalizacija, so zadovoljivi za pogoje, v katerih so bili podatki zajeti. Zaradi amaterskih posnetkov, uporabe ene samcate kamere ter slabe postavitve pri preslikavah igralcev prihaja do odstopanj. V popolnih delovnih pogojih, kjer bi bilo več kamer, ki bi bile postavljene na boljših položajih, bi bili rezultati nekoliko točnejši, vendar nikoli dovolj precizni, zaradi nepopolne transformacije iz tridimenzionalnega v dvodimenzionalni prostor.

Keywords

računalniški vid;algoritmi;lokalizacija;rokomet;visokošolski strokovni študij;Aplikativna elektrotehnika;diplomske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UL FE - Faculty of Electrical Engineering
Publisher: [J. Mlakar]
UDC: 004.93.021(043.2)(0.034.2)
COBISS: 199590403 Link will open in a new window
Views: 62
Downloads: 25
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Handball player localisation in handball using deep pre-trained models
Secondary abstract: Artificial intelligence and computer vision are very useful and high-quality tools in sport, as they reduce the time needed to search, collect and analyse data. However, as artificial intelligence is not yet advanced enough for slightly advanced programs and ideas, human knowledge is currently still at a great advantage. If we imagine a program that only needed to take a snapshot of a handball match and instantly return segmented data of all the individuals and their shots, perhaps the design for such a program does not seem so complex. When we look into the design in more detail, we realise that many different approaches and work are needed to create such an application. One of the biggest basics for such an application would be the localization of players with deeply learned algorithms. Such localization relies heavily on human assistance, as the algorithms need to be physically launched. But we have to start somewhere, because that is the only way to get to the desired goals. With the help of various pre-prepared algorithms and our own knowledge, we can extract the location of players on the layout of a handball court from amateur footage of a handball match, or from three-dimensional to two-dimensional space. The results returned by the localisation are acceptable for the conditions in which the data were captured. Due to amateur shots, the use of a single camera and poor placement, there are deviations in the mapping of the players. In perfect working conditions, where there would be several cameras placed in better positions, the results would be slightly more accurate, but never precise enough, due to the incomplete transformation from three-dimensional to two-dimensional space.
Secondary keywords: Computer vision;algorithm;localization;handball;
Type (COBISS): Bachelor thesis/paper
Study programme: 1000315
Thesis comment: Univ. v Ljubljani, Fak. za elektrotehniko
Pages: 1 spletni vir (1 datoteka PDF (IX, 43 str.))
ID: 24398902