Dense stereoscopic reconstruction with semi-global matching

Ovaj rad razmatra metode guste stereoskopske rekonstrukcije na standardnim skupovima slika. Cilj takvih metoda je odrediti oblak točaka koji najbolje odgovara prostoru projiciranom na slici. Poseban naglasak je na metodama koje favoriziraju glatku rekonstrukciju. Kao glavni algoritam uvedeno je poluglobalno podudaranje koje se temelji na pretpostavci da radimo uglavnom s glatkim površinama. Zato algoritam kažnjava svaki prekid u dubini slike. Kao ulaze prima cijene izračunate u metodama lokalne korespondencije. Algoritmi su evaluirani na standardnim skupovima slika Middlebury i KITTI. Rezultati i usporedbe metoda su prikazani grafički i komentarima.This thesis considers methods for dense stereoscopic reconstruction on standard image datasets. The aim of such methods is to determine the cloud of points that best suits the space projected onto the picture. Particular emphasis is placed on methods that favor smooth reconstruction. As the main algorithm, the thesis introduces semi-global matching which is based on the assumptions that we are mainly concerned about smooth surfaces. That is the reason why the algorithm penalizes every discontinuity in the depth of the image. It accepts, as input, the costs calculated by local correspondence methods. Algorithms are evaluated on standard image datasets Middlebury and KITTI. Results and comparisons of the methods are shown graphically and discussed