Creating a 3D Model of Indoor Environments by Fusion of 3D Point Clouds Acquired by a 3D Camera

U radu je objašnjen algoritam za stvaranje 3D modela prostorije fuzijom 3D oblaka točaka. Korištenjem navedenog algoritma, koji je implementiran unutar RVL (Robot Vision Library) biblioteke, implementirana je simulacija mjernog šuma koji se javlja u stvarnim mjerenjima s RGB-D kamerama. Algoritam radi na principu aktivne vizije koja se temelji na dva kriterija: što veći postotak još neistražene scene mora biti sadržan u sljedećem pogledu i dio scene koji je potpuno obuhvaćen sljedećim i jednim od prethodnih pogleda sadrži dovoljno informacije za uspješno poravnanje. Za izračunavanje korespondencije točaka u 3D oblaku točaka koristi se ICP algoritam koji ima tri uvjeta za ispravnu korespondenciju: euklidska udaljenost između dviju točaka mora biti manja od one uzrokovane rotacijom kamere za 3°, razlika normala tih točaka mora biti manja od 10° i razlika udaljenosti od kamere obje točke mora biti manja od 0,03m. Algoritam je testiran na šest različitih 3D modela prostorija preuzetih iz Replica baze. Dobiveni modeli uspoređeni su s originalnim. Maksimalna i srednja pogreška rotacije su prikazane za različite poglede kroz različite iteracije 3D modela prostorija iz Replica baze te su prikazani odgovarajući grafovi za njih.An algorithm for creating 3D room model by fusing 3D point clouds was explained in this paper. This algorithm, which is implemented as a part of RVL (Robot Vision Library), was used to simulate noise which occurs in real-time measurements with RGB-D cameras. The algorithm works on the principle of active vision which has two criteria: next view has to incorporate the largest possible percentage of yet unexplored scene and part of the scene that is completely incorporated in both the next and one of the previous views contains enough information for succesful registration. An ICP algorithm, which has three conditions for succesful registration, was used to calculate point correspondence in 3D point clouds. Those conditions are: Euclidean distance between two points has to be less than the point shift caused by camera rotation of 3°, difference in point normals of these points has to be less than 10° and difference in distances from the camera of these points has to be less than 0,03m. The algorithm was tested on six different 3D room models from Replica Dataset. Point cloud registration was tested and the created models were compared with the original ones. Maximal and average rotation errors were calculated for different views in different iterations of 3D room models from the Replica Dataset. These results were presented in the form of graphs for each 3D room model