Sohn, GunhoJahromi, Ali Baligh2020-05-112020-05-112019-082020-05-11https://hdl.handle.net/10315/37366In recent years, 3D indoor modeling has gained more attention due to its role in decision-making process of maintaining the status and managing the security of building indoor spaces. In this thesis, the problem of continuous indoor corridor space modeling has been tackled through two approaches. The first approach develops a modeling method based on middle-level perceptual organization. The second approach develops a visual Simultaneous Localisation and Mapping (SLAM) system with model-based loop closure. In the first approach, the image space was searched for a corridor layout that can be converted into a geometrically accurate 3D model. Manhattan rule assumption was adopted, and indoor corridor layout hypotheses were generated through a random rule-based intersection of image physical line segments and virtual rays of orthogonal vanishing points. Volumetric reasoning, correspondences to physical edges, orientation map and geometric context of an image are all considered for scoring layout hypotheses. This approach provides physically plausible solutions while facing objects or occlusions in a corridor scene. In the second approach, Layout SLAM is introduced. Layout SLAM performs camera localization while maps layout corners and normal point features in 3D space. Here, a new feature matching cost function was proposed considering both local and global context information. In addition, a rotation compensation variable makes Layout SLAM robust against cameras orientation errors accumulations. Moreover, layout model matching of keyframes insures accurate loop closures that prevent miss-association of newly visited landmarks to previously visited scene parts. The comparison of generated single image-based 3D models to ground truth models showed that average ratio differences in widths, heights and lengths were 1.8%, 3.7% and 19.2% respectively. Moreover, Layout SLAM performed with the maximum absolute trajectory error of 2.4m in position and 8.2 degree in orientation for approximately 318m path on RAWSEEDS data set. Loop closing was strongly performed for Layout SLAM and provided 3D indoor corridor layouts with less than 1.05m displacement errors in length and less than 20cm in width and height for approximately 315m path on York University data set. The proposed methods can successfully generate 3D indoor corridor models compared to their major counterpart.Author owns copyright, except where explicitly noted. Please contact the author directly with licensing requests.RoboticsElectronic Thesis or Dissertation2020-05-113D Indoor Corridor ReconstructionHypotheses VerificationOrientation MapGeometric ContextLayout SLAMModel-based MatchingLoop Closure