University of Twente Student Theses


Indoor Modelling Using RGB-D Data

Mehranfar, Mina (2013) Indoor Modelling Using RGB-D Data.

[img] PDF
Abstract:Various applications of indoor modelling in surveillance and emergency managements, navigations, positioning, robotics, forensics, virtual tours and so many on, brings increasing attentions to this field. Some of these applications need 3D models, e.g. a virtual tour is not possible in 2D models. Also 3D models interact with the users better and efficient. Meanwhile one of the recent technologies to acquire data for this use case, is using the RGB-D sensors. Kinect as one of these sensors, is the most famous and affordable cameras to capture the coloured point cloud. This research is proposed to develop a method for indoor 3D modelling using Kinect RGB-D data based on Manhattan-World assumptions. First we recognize the planes of indoor, using the characteristics of normal vectors of segments. Then using the grammar based modelling with a cube as the primitive shape, the 3D primitives are generated on all possible planes of indoor. In this way we first select the farthest segments of point cloud from the centroid of indoor, thus the first generated cube is the possible largest cube. Then we reconstruct new cubes using the remained segments and subtract them from the largest cube. In reconstructing the new cubes, we must preserve the orientation parameters of first cube to obey the Manhattan-World assumptions. The final model is written into VRML file and for better visualization of that, we add textures on the faces of model. The developed method is experimented on three different shaped datasets (L, U, X) and the results show that the method is valid for all of them. Also the method is implemented on real L-shaped point cloud with large registration errors. Despite of those error, all the existing planes are reconstructed in final model. The measurements on the reconstructed model is compared with the real measures to check the accuracy. On the planes with low registration error, we gain satisfactory accuracy. The developed method covers all the objectives of research, however there are some issues to improve the results. For example, we assumed to have only one plane of floor and ceiling without steps, that in future works, the reconstructed model is better to consider the steps in these two planes and also extending the method to model the scenes which don’t obey the Manhattan-World assumptions with non-perpendicular planes to each other is another follow up of this research. Key Words: Indoor Modelling, Kinect RGB-D data, Plane recognition, Grammar based modelling
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page