Implementation of Stereo Vision Engine
Translated title
Implementation af Stereo Vision Maskine
Author
Trillingsgaard, Tomas Brandt
Term
4. term
Education
Publication year
2016
Submitted on
2016-12-16
Pages
86
Abstract
Denne afhandling undersøger, hvordan man kan udvikle en hurtig og højpræcis stereovisionsmotor til implementering på en FPGA (et omprogrammerbart hardware, velegnet til parallel behandling). Stereovision estimerer dybde ved at sammenligne to billeder fra lidt forskellige vinkler, men området rummer både tekniske muligheder og praktiske udfordringer. To algoritmer — Efficient Edge Preserving Stereo Matching og Fast Cost-Volume Matching — er implementeret i Python og sammenlignet. Både beregningskompleksitet og kvaliteten af selve stereomatchingen er vurderet, hvilket fører til valget af Efficient Edge Preserving Stereo Matching som det bedste bud til videre FPGA-implementering. En endelig FPGA-implementering blev ikke færdiggjort, men arbejdet har identificeret og adresseret to centrale udfordringer: effektiv implementering af eksponentialfunktioner og håndtering af hukommelsesforbrug ved store billeder. For begge områder blev der fundet løsninger, som kan bane vej for en fremtidig implementering.
This thesis explores how to build a fast, high-precision stereo vision engine for implementation on an FPGA (a reconfigurable hardware platform suited to parallel processing). Stereo vision estimates depth by comparing two images taken from slightly different viewpoints, and the field involves both technical opportunities and practical hurdles. Two algorithms — Efficient Edge Preserving Stereo Matching and Fast Cost-Volume Matching — were implemented in Python and compared. We assessed computational complexity and the quality of the stereo matching, leading to the choice of Efficient Edge Preserving Stereo Matching for further FPGA implementation. A complete FPGA implementation was not achieved, but the work identified and addressed two key challenges: efficiently computing exponential functions and managing memory when processing large images. Solutions were found for both, providing a path toward future implementation.
[This abstract was generated with the help of AI]
Keywords
Documents