Growing 3D maps from 2D SLAM maps obtained by indoor SLAM with heterogeneous mobile robots


Thesis Type: Postgraduate

Institution Of The Thesis: Middle East Technical University, Faculty of Engineering, Department of Electrical and Electronics Engineering, Turkey

Approval Date: 2015

Thesis Language: English

Student: Burak Çetinkaya

Supervisor: AYDAN MÜŞERREF ERKMEN

Abstract:

This thesis focuses on growing 3D structure of an indoor environment from 2D maps constructed by Simultaneous Localization and Mapping (SLAM) algorithm at different height levels explored by heterogeneous vehicles. Vehicles differing mainly in their dynamics explore an indoor environment area using laser scanners at different height levels and creates 2D SLAM maps for each height level based on FastSLAM algorithm. This algorithm utilizes feature based scan matching and natural landmarks. These obtained 2D SLAM maps are then combined towards growing a 3D map in which three consecutive steps are followed namely; creating correspondences on the generated maps for correct alignment of the maps, generating correspondence points to use while connecting maps to each other and tiling the points which is the extrapolation method used between maps. Correspondences are to be established in the most adequate manner between the maps using topological adjacency relations via extracted key features from the maps so that these features can be matched and proper alignment between the maps can be established. Subsequently, regions that show similarities and dissimilarities between the maps are obtained based on the affinity of the grids belong to corresponding regions. Although, similar regions from different height levels are easy to grow from 2D to 3D by connecting the boundary regions as an extrapolation approach, for dissimilar regions, it is hard to find correspondence points to be used efficiently in an extrapolation. The novel approach presented in this thesis focuses on the generation of correspondence points and their extrapolation between SLAM maps. Tree structures are established inside the dissimilar regions and connected vertices on these tree structures are utilized as correspondence points. Extrapolation is generated by tiling that involves triangular meshing between generated correspondences between maps which eventually become surfaces of the 3D structures completed 2D maps. The thesis concludes with an intensive discussion of actual implementation on hardware followed by a full sensitivity analysis of the 3D map growing system proposed.