Ligikaudne siseruumi mõõtmete hindamine kasutades liikumisest struktuuri eraldamise võtteid
Date
2017
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Tänapäeva uusimaid võtteid kasutades on võimalik liikumisest piltide peal eraldada struktuure, mille täpsust võib võrrelda laser skännerite täpsusega. Kahjuks vajavad need algoritmid siiski huvipunkte ning tekstuurseid pindasid, et anda parimaid tulemusi. Enamus algoritmidel on halb jõudlus kui nad peavad töötlema monotoonsete värvidega või tekstuurivaeseid pindu. Lisaks on tihti algoritmide tulemustes auke, mida algses stseenis ei esine. Selline projektsioon ei ole väga kasulik ega kasutatav juhul kui on stseenist oodatakse järjepidavust ja täielikkust, mitte detailide täpsust. Käesolevas töös proovib autor kasutada liikumisest struktuuri eraldamise võtteid koos uudsete ideedega, selleks et luua projektsioon siseruumist fookusega toa olemusele (õiged pikkuste suhted, õige põranda plaan), mitte objektide detailsusele. Töö eesmärgiks on välja pakkuda algoritm, mis suudab taasluua stseeni, mida oleks võimalik kasutada rakendustes, kus rõhk on stseeni täielikkusel, mitte detailsusel (nt. õige tee leidmise probleemid).
Nowadays structure from motion algorithms have become accurate enough to compete with laser scanner accuracy, however most of the algorithms require points of interest and textured surfaces in order to give better results. Most algorithms will have poor performance when it comes to monotonically coloured or textureless surfaces. Furthermore, the output of the algorithms will have gaps in the projection of the structure it is trying to recreate. This kind of projection would be useless in a case where consistency and completeness of surfaces is more important than the level of detail. In this thesis the author will try to use structure from motion techniques and new ideas to create a projection of an interior room which focuses on the essence of the room (I.e aspect ratio, correct floor plan) rather than on the level of detail of objects in the room. The goal of this thesis will be to create an algorithm which can generate a projection out of a sparse point cloud (result of SfM) that is consistent enough to allow it to be used for applications that require a more complete model rather than a detailed one (I.e robot pathfinding, indoor people tracking).
Nowadays structure from motion algorithms have become accurate enough to compete with laser scanner accuracy, however most of the algorithms require points of interest and textured surfaces in order to give better results. Most algorithms will have poor performance when it comes to monotonically coloured or textureless surfaces. Furthermore, the output of the algorithms will have gaps in the projection of the structure it is trying to recreate. This kind of projection would be useless in a case where consistency and completeness of surfaces is more important than the level of detail. In this thesis the author will try to use structure from motion techniques and new ideas to create a projection of an interior room which focuses on the essence of the room (I.e aspect ratio, correct floor plan) rather than on the level of detail of objects in the room. The goal of this thesis will be to create an algorithm which can generate a projection out of a sparse point cloud (result of SfM) that is consistent enough to allow it to be used for applications that require a more complete model rather than a detailed one (I.e robot pathfinding, indoor people tracking).