State estimation and multiple moving object tracking on Riemannian manifolds

doctoral thesis

• Title (croatian): Estimacija stanja i praćenje više gibajućih objekata na Riemannovim mnogostrukostima
• Author: Borna Bićanić
• Mentor: Ivan Petrović (mentor)
• Committee member: Ivan Marković (predsjednik povjerenstva)
• Committee member: Mario Bukal (član povjerenstva)
• Committee member: Robert Cupec (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; (Department of Electronics, Microelectronics, Computer and Intelligent Systems); Zagreb
• Defense date and country: 2023-07-17, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Universal decimal classification (UDC): 004 - Computer science and technology. Computing. Data processing

Abstract

This thesis studies the problem of state estimation and multi-target tracking. First, the thesis presents a novel method for pedestrian tracking from the sequence of images. In visual tracking, it is also possible to obtain the appearance features by utilizing deep neural networks. However, fusing appearance cues represented by deep features of the neural networks together with the kinematic cues inside a classical multi-target tracking framework proves to be a non-trivial problem due to the nonlinearity of the underlying space of deep embeddings. To this end, this dissertation utilizes the tools of Riemannian geometry to tackle this problem. This thesis proposes a novel filtering method based on the unscented transform for systems whose state-space is a tangent bundle of a Riemannian manifold. By modeling the state-space as a tangent bundle of a manifold, it is possible to simultaneously estimate the position of the object on the manifold and its velocity, which lies in the tangent space of the manifold. In other words, it allows the use of the constant velocity motion model for state estimation on Riemannian manifolds. Although this can also be achieved with the Riemannian manifold extended Kalman filter, the proposed method avoids the tedious calculation of Jacobians. In the final part of the thesis, the state estimation methods for Riemannian manifolds are adapted to the multi-target case resulting in the novel multi-target tracking method for Riemannian manifolds based on probabilistic data association.

Abstract (croatian)

Ovaj doktorat proučava problem estimacije stanja i praćenja više objekata. Prvo, rad predstavlja novu metodu za praćenje pješaka iz niza slika. U vizualnom praćenju također je moguće dobiti značajke izgleda korištenjem dubokih neuronskih mreža. Međutim, fuzija znakova izgleda predstavljenih dubokim značajkama neuronskih mreža zajedno s kinematičkim značajkama unutar klasičnih algoritama za praćenje više objekata pokazalo se netrivijalnim problemom zbog nelinearnosti prostora dubokih značajki. U tu svrhu, ova disertacija koristi alate Riemannove geometrije za rješavanje ovog problema. Ova disertacija predlaže novu metodu filtriranja koja se temelji na transformaciji temeljenoj na sigma točkama za sustave čiji je prostor stanja tangencijalni snop Riemannove mnogostrukosti. Modeliranjem prostora stanja kao tangencijalnog snopa mnogostrukosti, moguće je istovremeno procijeniti položaj objekta na mnogostrukosti kao i njegovu brzinu, koja leži u tangencijalnom prostoru mnogostrukosti. Drugim riječima, omogućuje korištenje modela gibanja konstantne brzine za procjenu stanja na Riemannovim mnogostrukostima. Iako se to također može postići s proširenim Kalmanovim filtrom za Riemannove mnogostrukosti, predložena metoda izbjegava zamorno izračunavanje Jakobijana. U završnom dijelu rada, metode estimacije stanja za Riemannove mnogostrukosti proširene su na problem praćenja više objekata, što je rezultiralo novom metodom praćenja više objekata za Riemannove mnogostrukosti koja se temelji na vjerojatnosnoj asocijaciji podataka.

• Keywords: state estimation; multi-target tracking; Bayesian estimation; Kalman filter; Riemannian geometry; nonlinear estimation
• Keywords (croatian): estimacija stanja; praćenje više objekata; Bayesova estimacija; Kalmanov filter; Riemannova geometrija; nelinearna estimacija
• Language: english
• URN:NBN: urn:nbn:hr:168:573863
• Promotion: 2023
• Study programme: Title: Computer Science; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti (Doktor znanosti)
• Catalog URL: https://lib.fer.hr/cgi-bin/koha/opac-detail.pl?biblionumber=54314
• Type of resource: Text
• Extent: xvii, 123 str.
• File origin: Born digital
• Access conditions: Open access
• Created on: 2024-02-09 11:18:53