Learning-based 3D scene reconstruction using RGBD cameras

Abstract

Machine learning methods and object recognition algorithms have improved much in the past decade, but computer perception and object recognition remain some of the biggest challenges of modern engineering. A close relative of these is scene reconstruction, in which the computer attempts to create a digital reconstruction of the environment it is perceiving. This thesis considers the use of RGB-D cameras in room reconstruction, which is a particularly interesting field of scene reconstruction for mobile robots. RGB-D cameras and reconstruction algorithms are not as widespread as LiDAR-based applications in robots but can be cost-efficient replacements in certain situations. In this thesis, applications of room reconstruction methods are also discussed by giving an overlook of state-of-the-art algorithms. This thesis is divided into two parts. First in the literary review section elucidates upon the basic theory of the subject and presents the current state of room reconstruction. The experi-mental part of the work examines in detail the operation of the algorithms used. Finally, achieved results are displayed and analyzed along with potential future research. The reconstruction of the test environment was manufactured with a single moving RGB-D capable camera, and object recognition semantics was applied to this scene. This was achieved by applying InstanceFusion on a dataset collected with a Stereolabs ZED camera. The achieved reconstruction is not as good as examples in research material, and reasons for this are explored.