Computation of disparity in stereo images using three-dimensional recursive search

doctoral thesis

• Title (croatian): Izračun dispariteta u stereoslikama s pomoću trodimenzionalnoga rekurzivnoga pretraživanja
• Author: Miroslav Rožić
• Mentor: Tomislav Pribanić (mentor)
• Committee member: Tomislav Pribanić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; (Department of Electronic Systems and Information Processing); Zagreb
• Defense date and country: 2019, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Electrical Engineering; Automation and Robotics
• Universal decimal classification (UDC): 621.3 - Electrical engineering

Abstract

Disparity computation is a crucial step in the process of stereo 3D reconstruction, which has been a heavily investigated topic and one of the important problems in the field of computer vision. Many methods and approaches have been devised to provide a dense disparity map for a stereo image pair, with approaches roughly divided into local methods, which determine the correct disparity by choosing the value with an optimum matching cost in a local region, or global methods, which attempt to optimize a global energy equation for the entire scene. The global methods, while generally yielding better results, are more computationally intensive and are generally not suitable for real-time or embedded applications. The aim of this research was to propose new methods for disparity computation which would yield greater accuracy of the computed disparity map and faster computation by combining existing stereo algorithms with the three-dimensional recursive search algorithm (3DRS). The 3DRS algorithm was originally devised for motion estimation, frame rate up-conversion and de-interlacing in high definition televisions, and it can generate a coarse correspondence map (depending on the use, either the optical flow or disparity map) within a very short amount of time. In this dissertation, two methods based on the combination of the 3DRS algorithm with the local Winner-take-all method or the global Dynamic Programming method are presented. The proposed methods are evaluated using the Middlebury image set. It is shown that the proposed hybrid methods significantly outperform the original methods in execution time while maintaining or improving upon the accuracy of both methods. In addition, the hybrid methods are shown to exhibit greater robustness by reducing the dependency on specific external parameters. Scientific contribution: • A robust local method for computing stereo disparity maps initialized by the three-dimensional recursive search algorithm. • A robust global method for computing stereo disparity maps initialized by the three-dimensional recursive search algorithm. • A method for comparative analysis of the developed methods and the reference global and local stereo matching methods.

Abstract (croatian)

Izračun dispariteta je ključan korak u postupku stereo 3D rekonstrukcije, koje je već dugo istraživana tema i jedan od važnih problema u polju računalnog vida. Mnoge su metode i pristupi predloženi sa svrhom dobivanja guste mape dispariteta za par stereo slika, koje ugrubo možemo podijeliti na lokalne metode, koje utvrđuju ispravni disparitet odabirom vrijednosti sa optimalnom cijenom uparivanja u lokalnoj regiji slike, ili globalne metode, koje pokušavaju optimirati globalnu energiju za sve piksele u sceni odjednom. Iako postižu bolje rezultate dispariteta, globalne su metode uglavnom iznimno računalno zahtjevne te nisu podesne za primjene u ugradbenim sustavima, odnosno za rad u stvarnom vremenu. Cilj je ovog istraživanja bio predložiti nove metode izračuna dispariteta sa poboljšanom točnošću i ubrzanim izvršavanjem, dobivene kombiniranjem postojećih algoritama sa algoritmom trodimenzionalnog rekurzivnog pretraživanja (3DRS). Algoritam 3DRS je originalno zamišljen za procjenu pokreta, interpolaciju slika sa ciljem povećanja frekvencije video signala, te rasplitanja video signala u televizijskim sustavima visoke definicije, te može generirati grubu mapu korespondencija (ovisno u slučaju korištenja, optički tok ili mapu dispariteta) za par slika u veoma kratkom vremenu. U ovom su radu predstavljene dvije metode temeljene na kombinaciji algoritma 3DRS sa lokalnom metodom „pobjednik-uzima-sve“ te globalnom metodom temeljenom na postupku dinamičkog programiranja. Predložene su metode evaluirane na „Middlebury“ skupu slika. Pokazano je da predložene hibridne metode značajno nadmašuju originalne metode u vremenu izvođenja, zadržavajući pritom ili unaprjeđujući točnost originalnih metoda. Dodatno, hibridne metode pokazuju veću robusnost od originalnih, smanjujući ovisnost kvalitete izlaznih rezultata o ulaznim parametrima. Znanstveni doprinos: • Robusna lokalna metoda za izračun mapa stereo dispariteta inicijalizirana algoritmom 3D rekurzivne pretrage. • Robusna globalna metoda za izračun mapa stereo dispariteta inicijalizirana algoritmom 3D rekurzivne pretrage. • Metoda za komparativnu analizu razvijenih metoda i referentnih globalnih i lokalnih metoda stereo uparivanja

• Keywords: Stereo matching; disparity; 3DRS; Hybrid method; Winner-take-all; Dynamic programming
• Keywords (croatian): Stereo uparivanje; disparitet; 3DRS; Hibridna metoda; Pobjednik-uzima-sve; Dinamičko programiranje
• Language: english
• URN:NBN: urn:nbn:hr:168:393325
• Study programme: Title: Electrical Engineering and Computing; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti elektrotehnike i računarstva (Doktor znanosti elektrotehnike i računarstva)
• Type of resource: Text
• Extent: 87 str. ; 30 cm
• File origin: Born digital
• Access conditions: Open access
• Created on: 2020-01-22 06:51:38