prikaz prve stranice dokumenta Stvarnovremenska detekcija objekata u prometu primjenom tehnika strojnog učenja
Access restricted to students and staff of home institution
master's thesis
Stvarnovremenska detekcija objekata u prometu primjenom tehnika strojnog učenja
2019. urn:nbn:hr:190:311309
Jukić, Dino
University of Rijeka
Faculty of Engineering
Department of Computer Engineering
Section of Intelligent Computing Systems

Cite this document

Jukić, D. (2019). Stvarnovremenska detekcija objekata u prometu primjenom tehnika strojnog učenja (Master's thesis). Rijeka: University of Rijeka, Faculty of Engineering. Retrieved from https://urn.nsk.hr/urn:nbn:hr:190:311309

Jukić, Dino. "Stvarnovremenska detekcija objekata u prometu primjenom tehnika strojnog učenja." Master's thesis, University of Rijeka, Faculty of Engineering, 2019. https://urn.nsk.hr/urn:nbn:hr:190:311309

Jukić, Dino. "Stvarnovremenska detekcija objekata u prometu primjenom tehnika strojnog učenja." Master's thesis, University of Rijeka, Faculty of Engineering, 2019. https://urn.nsk.hr/urn:nbn:hr:190:311309

Jukić, D. (2019). 'Stvarnovremenska detekcija objekata u prometu primjenom tehnika strojnog učenja', Master's thesis, University of Rijeka, Faculty of Engineering, accessed 27 November 2024, https://urn.nsk.hr/urn:nbn:hr:190:311309

Jukić D. Stvarnovremenska detekcija objekata u prometu primjenom tehnika strojnog učenja [Master's thesis]. Rijeka: University of Rijeka, Faculty of Engineering; 2019 [cited 2024 November 27] Available at: https://urn.nsk.hr/urn:nbn:hr:190:311309

D. Jukić, "Stvarnovremenska detekcija objekata u prometu primjenom tehnika strojnog učenja", Master's thesis, University of Rijeka, Faculty of Engineering, Rijeka, 2019. Available at: https://urn.nsk.hr/urn:nbn:hr:190:311309

Please login to the repository to save this object to your list.
Metadata
TitleStvarnovremenska detekcija objekata u prometu primjenom tehnika strojnog učenja
Title (english)Real-time Traffic-Object Detection using Machine Learning
AuthorDino Jukić
MentorIvan Štajduhar (mentor)
Committee memberJonatan Lerga (predsjednik povjerenstva)
Committee memberGoran Mauša (član povjerenstva)
Committee memberIvan Štajduhar (član povjerenstva)
GranterUniversity of Rijeka
Faculty of Engineering
(Department of Computer Engineering)
(Section of Intelligent Computing Systems)
Rijeka
Defense date and country2019-11-21, Croatia
Scientific / art field,
discipline and subdiscipline		TECHNICAL SCIENCES
Computing
Abstract
Cilj ovog diplomskog rada bilo je proučiti razne modele koji se mogu koristiti za detekciju objekata u prometu u statičkom i u dinamičko-stvarnovremenskom okruženju te napraviti vlastiti detektor za predviđanje klasa uz pomoć metoda strojnog učenja. Motivacija ovog rada je proizašla iz ideje stvaranja osnove za autonomnu vožnju, pošto se u današnje vrijeme strojno učenje primjenjuje u gotovo sve aspekte naših života, pa tako i u automobilskoj industriji, rad ove vrste se pokazao kao značajnim i korisnim. Kako bi čitatelju tema bila jasnija, ukratko smo se upoznali s osnovnim pojmovima koji predstavljaju sastavni dio strojnog učenja te smo predstavili i objasnili razlike između pojedinih metoda – klasifikacija i lokalizacija – koje se koriste za prepoznavanje objekata. Sukladno tome, u radu smo se dotakli teme prepoznavanja objekata i raznih modela kojima je to omogućeno te je za svaki model opisan način njihova rada, opisane su njihove prednosti i nedostaci. Model koji se pokazao najzanimljivijim te koji se koristio u svrhe ostvarenja zadatka diplomskog rada je bio Faster RCNN. Inception V2 model se pokazao kao zlatnom sredinom u usporedbi s potrebama ovoga rada. Pristup rješavanju rada se sastojao od tri osnovna koraka: prikupljanju podataka, treniranju modela i testiranju modela. Cilj prikupljanja podataka je bilo sakupiti što veći broj raznovrsnih slika objekata koji se pojavljuju u prometu, što je u konačnici rezultiralo s 2000 različitih fotografija (svaka s različitim brojem objekata), s 10 različitih klasa objekata. Ideja je bilo koristiti vlastite skupove slika umjesto preuzetih online kako bi mogli bolje razumjeti podatke. Podaci su se sakupljali kroz dulji vremenski period te je bilo potrebno filtrirati skup fotografija i skalirati na manju veličinu kako bi se olakšalo treniranje. Za treniranje potrebno je bilo svaki objekt ručno označiti i zapisati te podesiti radno okruženje kako bi uspješno istrenirali model. Tijek treniranja se pokazao rutinskim procesom koji je zahtijevao ponavljanje postupka treniranje dokle nismo bili zadovoljni s rezultatima te u konačnici u svrhe boljeg testiranja koristila su se četiri pristupa: Dva testiranja u statičkom okruženju te dva u dinamičkom; osnova svrha testiranja je bilo istražiti efikasnost rada modela u odnosu na njegovu brzinu i preciznost. U konačnici, možemo reći da je detektor zadovoljio kriterije te se Faster RCNN dokazao kao dostatnim rješenjem koje ostavlja mogućnost za daljnje napredovanje, poboljšanje, dodatne implementacije u budućnosti.
Abstract (english)
The purpose of this thesis was to study various models that can be used to detect objects in traffic in a static and dynamic real-time environment and to create our own class prediction detector using machine learning. The motivation behind this thesis stands from the idea of creating the basis for autonomous driving, which in today's time of rigorous learning is appropriate in almost all aspects of life, including in the automotive industry, and work of this kind has proven to be both meaningful and useful. For the purpose of better understanding the topic, we briefly introduced ourselves to the basic concepts that are an integral part of machine learning and explained the differences between individual methods – classification and localization – which are used to identify objects. We have also familiarised ourselves with the topic of object detection accordingly and with the various models which make detection possible, and for each of them we described how they work, what their advantages and disadvantages are. The model that proved to be the most interesting, which was used for the purpose of writing our thesis, was the Faster RCNN model. The Inception V2 turned out to be the best overall solution for the purposes of our thesis. The approach of problem solving consisted of three steps: gathering data, tranining the model and testing out the model. The aim of collecting data was to acquire as many diverse images of objects in traffic, which ultimately resulted in 2000 different images (each consisted of a different number of objects), with 10 unique classes of objects. The idea was to use our provided sets of images for the purpose of better understanding and handling the data. The data was collected over a longer period of time and was necessary to filter and resize the photos to improve the tranining. For the training, each individual object had to be manually labelled and the work environment needed to be adjusted. The training proved to be a routine process that required repetition until we were satisfied with the results, and in the end, for the purposes of better testing, four different methods were used: Two in a static environment and two in a dynamic; the purpose of it was to study the efficiency of the model with respect to its speed and precision. Ultimately, we can say that the detector met the criteria and the Faster RCNN has proven to be an excellent solution that leaves room for further advancment, improvement and further implementation in the future.
Keywords
Keywords (english)
Languagecroatian
URN:NBNurn:nbn:hr:190:311309
Study programmeTitle: Computer Engineering
Study programme type: university
Study level: graduate
Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
Type of resourceText
File originBorn digital
Access conditionsAccess restricted to students and staff of home institution
Terms of useLicence In copyright icon
Created on2020-07-27 10:17:13