Detekcija igrača u snimkama nogometnih utakmica temeljena na dubokom učenju

master's thesis

• Title (english): Player Detection in Football Game Video Based on Deep Learning
• Author: Ivan Matak
• Mentor: Zoran Kalafatić (mentor)
• Committee member: Zoran Kalafatić (predsjednik povjerenstva)
• Committee member: Tomislav Hrkać (član povjerenstva)
• Committee member: Alan Jović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2019-07-11, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Rad se bavi problemom detekcije i praćenja nogometnih igrača na utakmici. Za detekciju se koristi YOLOv3 arhitektura te je pobliže objašnjena njena funkcija gubitka, način detekcije objekata, predikcija omeđujućeg pravokutnika (eng. bounding box) te način klasifikacije objekata. Za praćenje se koristi heuristika koja uz pomoć modela zasnovanog na izgledu (eng. appearance model), modela pokreta i algoritma za klasifikaciju timova dodjeljuje identifikatore igračima te ažurira pozicije igrača tijekom utakmice. Opisani su problemi praćenja te način na koji su riješeni. Uz to, opisan je način učenja klasifikatora za timove te način treniranja i testiranja sustava za detekciju. Skup podataka koji je korišten za razvoj oba sustava je nogometna utakmica snimana nepomičnom kamerom između GNK Dinamo Zagreb i RNK Split.

Abstract (english)

This thesis deals with the problem of detection and tracking football players in the game video. For detection, YOLOv3 architecture is used. Its loss function, way of detecting objects, its prediction of bounding boxes and way of classifying objects are explained in more detail in the thesis. The tracking uses a heuristics that assigns identifiers to players and updates their positions during the match with the help of appearance model, motion model and classification algorithm (kNN). Also, the problems that tracking system encounters during its execution are described along with the way how that problems were solved. In addition, it is described how was kNN classifier trained and how was the object detection system trained and tested. The dataset used to develop both systems was a football game video recorded by a stationary camera between GNK Dinamo Zagreb and RNK Split.

• Keywords: praćenje objekata; detekcija objekata; YOLOv3 arhitektura; heuristike; računalni vid; duboko učenje; strojno učenje; PyTorch
• Keywords (english): object tracking; object detection; YOLOv3 architecture; heuristics; computer vision; deep learning; machine learning; PyTorch
• Language: croatian
• URN:NBN: urn:nbn:hr:168:153168
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2020-01-07 20:10:59