Vizualno sledenje objektov na vgrajenih napravah

diplomsko delo

Nik Prinčič (Author), Luka Čehovin (Mentor)

Abstract

V okviru diplomskega dela je bilo implementirano in ovrednoteno delovanje vizualnega sledilnika na vgrajeni napravi Luxonis OAK-1. Izbran je bil sledilnik STARK, ki spada v družino sledilnikov, ki jih sestavljajo globoke nevronske mreže. Bolj specifično sledilnik temelji na arhitekturi transformer, ki je trenutno uporabljena v vseh najboljših vizualnih sledilnikih. Sledilnik je bilo potrebno prilagoditi ter ga prevesti v OpenVINO format, ki omogoča uporabo na vgrajeni napravi. Poleg tega je bilo potrebno zasnovati cevovod, po katerem se podatki na napravi pretakajo. Vgrajena naprava tako izvaja vse potrebne operacije od zajema slik do napovedi pozicije brez vpletenosti gostiteljskega sistema. S tem smo dosegli, da so performance sledenja neodvisne od gostiteljskega sistema. Vse verzije sledilnika smo eksperimentalno ovrednotili na testih VOT2021 in VOT2022 in s tem dokazali, da se ob prenosu na vgrajeno napravo, delovanje sledilnika ni poslabšalo.

Keywords

računalniški vid na vgrajenih napravah;DepthAI;vizualni sledilnik;vizualno sledenje objektov;visokošolski strokovni študij;diplomske naloge;

Data

Language:	Slovenian
Year of publishing:	2023
Typology:	2.11 - Undergraduate Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[N. Prinčič]
UDC:	004.93(043.2)
COBISS:	147286787
Views:	31
Downloads:	5
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Visual object tracking on embedded devices
Secondary abstract:	In this diploma thesis, a state-of-the-art visual tracker is implemented and evaluated on the embedded device Luxonis OAK-1. The tracker STARK is chosen, which belongs to the family of deep neural network-based trackers. More specifically the tracker uses the novel transformer neural network architecture, which is making its way into increasingly more best performing visual trackers. During the implementation process we had to modify the tracker, convert it to the OpenVINO format, which can be used for inference on the embedded device. To run the processing completely on the embedded device the correct pipeline architecture was also developed, which allows for all the processing to be executed on the embedded device and consequently allows autonomous operation of the embedded device. With this level of autonomy, we successfully decoupled tracking performance from the performance of the host system. We evaluated all versions of the tracker on the VOT2021 and VOT2022. With the evaluation we proved that the tracker ported to the embedded device doesn't lose any tracking performance.
Secondary keywords:	embedded computer vision;DepthAI;visual tracker;computer science;diploma;Računalniški vid;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS):	Bachelor thesis/paper
Study programme:	1000470
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	45 str.
ID:	18326731