Razvoj postopka za avtomatsko ocenjevanje sloga smučarskih skokov

Štepec, Dejan (author), Skočaj, Danijel (mentor)

Smučarski skoki so med Slovenci zelo priljubljen šport, predvsem zaradi uspehov naših športnikov. V magistrskem delu razvijemo metodo za avtomatsko ocenjevanje sloga smučarskih skokov iz videoposnetkov. Kot glavni vir za napoved sodniških ocen uporabimo razporeditev delov telesa in smuči skozi let smučarskega skakalca. Obstoječo metodo za detekcijo delov telesa uporabimo na domeni smučarskih skokov s pomočjo posebne zbirke podatkov, ki jo zgradimo v ta namen. Metodo za detekcijo delov telesa tudi ustrezno spremenimo, da omogoča detekcijo delov smuči. Detektirani deli telesa in smuči tvorijo vhod v metodo za napoved sodniških ocen, kjer uporabimo arhitekturo konvolucijskih nevronskih mrež na časovno odvisnih podatkih. Tako naučen model ima napako napovedi, ki je konkurenčna pravim sodnikom.

• Language: Slovenian
• Keywords: smučarski skoki, sodniške ocene, konvolucijske nevronske mreže, detekcija delov telesa
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2017
• PID: 20.500.12556/RUL-96714
• Publication date in RUL: 12.10.2017

Secondary language

• Language: English
• Title: Development of an approach for automatic ski jump style scoring

Abstract

Ski jumping has always been a very popular sport in Slovenia, mostly due to success of our sportsmen. The goal of of this master's thesis is to develop a method for automatic ski jump scoring from videos. As our main source of information we use locations of human body parts along with skis to capture a full body movement of the entire ski jump. We have used an existing method for human pose estimation from images on the domain of ski jumping with the help of specially built dataset. We extend the method for human pose estimation with the support for ski parts detection. Combined locations of human body parts and ski parts represents an input for the method that performs scoring of the ski jump style. The approach is based on convolutional neural networks that are atypically used on a time series data. Our method is able to operate with an error comparable to real judges.

• Keywords: ski jumping, ski jump scoring, convolutional neural networks, human pose estimation