Analysis of selected camera-based structure from motion algorithms

Poustvarjanje replike objekta iz delno prekrivajočih se fotografij je cilj ocenjevanja strukture iz gibanja. Za izvedbo popolne rekonstrukcije, je potrebno implementirati algoritme posamezne stopnje cevovoda. V iskanju učinkovitih in zanesljivih tehnik, smo zbrali raznolike algoritme na področjih ekstrakcije značilnic, opisa značilnic, ujemanja značilnic, ocenjevanja pozicije in rotacije, ekstrakcije oblaka točk ter 3D rekonstrukcije. Algoritme posamezne stopnje cevovoda smo nato primerjali po učinkovitosti, hitrosti ter zanesljivosti. Metode predstavljene v diplomskem delu so testirane na umerjenem setu rekonstruiranih 3D objektov. Podobnost med končno rekonstrukcijo ter zajetim 3D objektom je izmerjena z naborom orodij MeshLab. Kot najboljša se je izkazala kombinacija algoritmov: SURF za zaznavo značilnic, Približno iskanje bližnjih sosedov in PROSAC za iskanje pravilnih korespondenc ter ocenjevanje fundamentalne matrike, Inkrementalna struktura iz gibanja za ocenjevanje pozicij in rotacij kamer ter Gosto stereo ujemanje za triangulacijo oblaka točk.Estimating a 3D representation of a real-world object using partially overlapping images is the main goal of a photogrammetry technique known as Structure from Motion. Whereas the structure from motion pipeline has been well established, the algorithms in practice vary significantly in the implementation of the different stages. In search of effective and reliable techniques we have described several algorithms for feature detection, feature description, feature matching, pose estimation, point cloud extraction and 3D reconstruction. Each algorithm has been evaluated in terms of efficiency, speed and robustness. Methods presented in this thesis are tested on a dataset consisting of reconstructed 3D models and consecutive images of objects. For each model similarity between the reconstructed mesh and the original 3D model has been measured using the MeshLab toolbox. In summary the best performing methods in the reconstruction pipeline proved to be: SURF for feature detection, Fast approximate nearest-neighbour search for feature matching, PROSAC for fundamental matrix estimation, Incremental Structure from Motion for camera pose estimation and Dense stero matching for point cloud reconstruction