Discriminative correlation filter with segmentation and context for robust tracking

master's thesis

Ajda Lampe (Author), Matej Kristan (Mentor)

Abstract

Visual object tracking is an area in the field of computer vision, which has seen great popularity increase due to a large availability of video data. There are many different tracking tasks, such as multiple object tracking, long-term tracking and specialized trackers, expected to perform well in a very specific domain. In this work, we focus on online generic short-term single object tracking, which can be considered the base visual tracking task and can be adaptable to any of the previously mentioned tasks. We propose a new tracker, based on correlation filtering, augmented with context information and a predicted object segmentation mask. The results on benchmarks fall far behind the current state-of-the-art, however the proposed method consistently outperforms baseline trackers, which shows the methods potential for future improvements.

Keywords

computer vision;visual object tracking;correlation;segmentation;computer science;computer and information science;master's thesis;

Data

Language:	English
Year of publishing:	2020
Typology:	2.09 - Master's Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[A. Lampe]
UDC:	004.93'1(043.2)
COBISS:	40159235
Views:	877
Downloads:	207
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	Slovenian
Secondary title:	Diskriminativni korelacijski filter s segmentacijo in uporabo konteksta za robustno sledenje
Secondary abstract:	Vizualno sledenje objektom je področje računalniškega vida, ki je v zadnjih letih doživelo velik razcvet, zahvaljujoč dostopnosti video vsebin. Problem lahko razdelimo na več podnalog, na primer sledenje več objektom, dolgoročno sledenje ali specializirano sledenje za točno določeno domeno. V tem delu se omejimo na splošne kratkoročne sledilnike, ki sledijo enemu objektu. To lahko namreč razumemo kot najbolj osnovno nalogo vizualnega sledenja, ki jo lahko razširimo za delovanje na prej omenjenih problemih. V delu predstavimo nov sledilnik, ki temelji na sledenju s korelacijskimi filtri, razširimo pa ga z uporabo kontekstne informacije in segmentacijske maske. V primerjavi z ostalimi sledilniki predlagana metoda sicer ne dosega rezultatov, primerljivih z najmodernejšimi sledilniki, vendar pa dosledno dosega boljše rezultate od osnovnejših sledilnikov, kar kaže na potencial metode za nadaljnje izboljšave.
Secondary keywords:	računalniški vid;vizualno sledenje objektom;korelacija;segmentacija;računalništvo;računalništvo in informatika;magisteriji;
Type (COBISS):	Master's thesis/paper
Study programme:	1000471
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	VIII, 76 str.
ID:	12149298