Učinkovita implementacija sledilnika CSR-DCF

MUHIČ, ANDREJ (author), Kristan, Matej (mentor)

V diplomski nalogi smo predstavili problem vizualnega sledenja poljubnemu objektu skozi sekvenco slik. Opisali smo diskriminativni korelacijski filter, ki predstavlja temeljno orodje za izgradnjo našega sledilnika. Osredotočili smo se na metodo CSR-DCF, ki korelacijske filtre izboljša z uporabo maske in kanalskih uteži. Maska določa, kateri piksli so pomembni za učenje filtra. Izračuna se jo na podlagi barvnega ujemanja ter prostorske zanesljivosti, uteži pa določajo diskriminativno moč posameznega kanala. Sledilnik je na lestvici VOT med najboljšimi, vendar je njegova glavna pomanjkljivost neučinkovita implementacija v jeziku Matlab. Ker smo želeli, da bi bil sledilnik široko dostopen ter da bi omogočal boljšo izrabo procesorske moči, smo ga implementirali v programskem jeziku C++ in ga vključili v prosto dostopno knjižnico OpenCV. S pomočjo ogrodja za testiranje kratkoročnih sledilnikov VOT smo izvedli analizo različnih parametrov sledilnika in primerjavo z originalno implementacijo v jeziku Matlab. Prišli smo do zaključka, da naša implementacija doseže primerljive rezultate ob višji hitrosti delovanja. Trenutno je sledilnik najboljši med vsemi sledilniki v knjižnici OpenCV, o čemer govori tudi dejstvo, da je dosegel prvo mesto v kategoriji realnočasovnih sledilnikov na lestvici VOT2017.

• Language: Slovenian
• Keywords: računalniški vid, vizualno sledenje, detekcija, korelacijski filtri, sledilnik
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2019
• PID: 20.500.12556/RUL-111522
• COBISS.SI-ID: 1538395075
• Publication date in RUL: 02.10.2019

Secondary language

• Language: English
• Title: An efficient implementation of the CSR-DCF tracker

Abstract

In this thesis we addressed the problem of visually tracking an object in a sequence of images. We described the discriminative correlation filter, which is a building block for our tracker. We also described the CSR-DCF method, that improves correlation filters by using a binary mask and channel weights. The mask determines which pixels are important for filter learning, and is calculated based on appearance likehood and spatial likehood, whereas channel weights determine the discriminative power of each channel. The tracker is one of the best on the VOT challenge, but its main drawback is the inefficient implementation in Matlab. We wanted to make the tracker widely accessible and have it run more efficiently. That is why we implemented it using C++ programming language and added it to OpenCV library. We used VOT framework to test the effects of different parameters on the speed and accuracy of this tracker, and to compare the original implementation in Matlab to ours. We concluded that our implementation produces similar results while running at a higher speed. The tracker is currently the best in OpenCV library, and was ranked first among realtime trackers on VOT2017.

• Keywords: computer vision, visual tracking, detection, correlation filters, tracker