bachelor thesis
Jovana Videnović (Author), Matej Kristan (Mentor), Alan Lukežič (Co-mentor)

Abstract

This thesis presents a novel approach for single object tracking in the presence of distractors. We define distractors as objects that are visually similar to the tracked target. This visual similarity in practice increases the tracking uncertainty, often leading to tracking failure when the tracker shifts from a target to a distractor. The proposed technique, named Multiple Hypotheses Single Object Tracker (MHSOT), builds upon a recent TransT tracker by including modules for motion estimation, feature extraction, and data association. Additionally, we gather a collection of sequences containing distractors to form the comprehensive Distractor dataset. We evaluate our method, along with the TransT, on the Distractor dataset and its short-term and long-term segments. MHSOT tracker achieves 51.7% mIoU on the whole Distractor dataset, surpassing the TransT by 3.6%. MHSOT and TransT perform better on short-term sequences than on long-term sequences. On short-term sequences, MHSOT outperforms the TransT by 6.1%, reaching 71.5% mIoU. On long-term sequences, MHSOT outperforms the TransT by 3.3% and achieves 50.1% mIoU. MHSOT demonstrates particular superiority over the reference method in re-detecting the target even after extended periods of absence and handling scenarios involving occlusion.

Keywords

computer vision;single object tracking;tracking-by-detection;data association;distractors;computer and information science;diploma thesis;

Data

Language: English
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [J. Videnović]
UDC: 004.93(043.2)
COBISS: 167529475 Link will open in a new window
Views: 46
Downloads: 9
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: Slovenian
Secondary title: Sledenje objekta ob prisotnosti distraktorjev
Secondary abstract: V diplomski nalogi predstavljamo inovativen pristop za obravnavo sledenja objekta ob prisotnosti distraktorjev. Distraktorje definiramo kot objekte, ki imajo veliko vizualnih podobnosti s tarčo. Ta vizualna podobnost v praksi poveča negotovost sledenja in pogosto vodi v neuspeh sledenja, ko sledilnik začne slediti distraktorju. Razvita metoda, poimenovana MHSOT, temelji na sledilniku TransT in uvaja nove module za oceno gibanja, izračun značilk ter asocijacijo podatkov. Poleg tega izberemo množico sekvenc, ki vključujejo distraktorje ter tako pridobimo celovit nabor podatkov imenovan Distractor dataset. Našo metodo, skupaj z referenčnim sledilnikom TransT, evalviramo tako na celotnem naboru podatkov, kot pa na njegovih kratkoročnih ter dolgoročnih sekvencah ločeno. Sledilnik MHSOT doseže 51.7% mIoU na množici Distractor dataset, kar predstavlja izboljšavo metode TransT za 3.6%. Sledilnika sta se bolje obnesla na kratkoročnih sekvencah, kjer je MHSOT dvignil uspešnost TransT metode za 6.1% mIoU, iz inicijalnih 67.4% na 75.1%. Na dolgoročnih sekvencah je MHSOT dosegel 50.1% mIoU oz. 3.3% mIoU več kot TransT. MHSOT izkazuje izjemno superiornost nad referečno metodo v redetekciji tarče, tudi po daljših obdobjih odsotnosti ter v primerih okluzije.
Secondary keywords: sledenje enega objekta;sledenje z zaznavanjem;asocijacija podatkov;distraktorji;univerzitetni študij;diplomske naloge;Računalniški vid;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS): Bachelor thesis/paper
Study programme: 1000468
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 74 str.
ID: 21439482