Algorithms for the Multiple Variants of Registration in 3D Range Data.

Abstract

Ever since the photography was born, there exists a high interest in the possibility of detecting a third dimension in the images obtained by a camera. This third dimension feature would allow the differentiation and easily filtering of the different objects present in the scene, and therefore to facilitate some of the main problematics in the computer vision research. In order to achieve this third dimension acquisition some techniques were historically applied, starting by the classical stereoscopy or other more current methods like Structure from Motion._x000D_ During last years, the evolution of the technology has made possible the appearance of devices which allows the direct retrieval of 3D information without the manual intervention of the user. The use of different light spectrums like laser or infrared light and their integration inside the camera case have allowed this important improvement, accompanied at the same time by a reduction of the components price which allows its use for the vast majority of the research community._x000D_ This thesis focuses on the main problems obtained in the use of the so-called range cameras, which, in addition to the usual RGB image, offers an additional image indicating the distance with respect to the objects in the scene. Thanks to the direct application of these range images, where each pixel corresponds to a distance, a 3D recreation of the observed scene can be directly obtained. One of the main advantages is that, if the camera is correctly calibrated, the 3D structure can be obtained with physical units of the real world (such as meters), and not with pixels._x000D_ Nevertheless, the use of these range cameras was not as ideal as supposed. Usually some captures of every object in the scene are needed in order to obtain a full reconstruction, and different materials could produce problems that interferes the correct position of the object. The presence of these inconveniences produces the necessity of using some algorithms to produce a correct final 3D structure._x000D_ The particularities of the 3D representation created from the range image become it appropriate to use as a support plate for placing other sources of information, like visible images or infrared images. The obtained 3D structure can be textured with these sources of information, giving an integrated result which could clarify some problems that cannot be solved by using the images separately. However, differences in the acquisition process between these types of images produce difficulties when they are fused._x000D_ In addition, in order to obtain a full-side representation of a scene usually some 3D captures from different points of views are required. This addresses to the so-called multiview registration problem, where it is necessary to identify the position and orientation of the range camera for each viewpoint in order to correctly join the corresponding 3D structures. Current technology devices like GPSs or IMUs could give this information, but usually is not accurate enough, so common visual elements between different range images must be detected in order to align them. This alignment is usually achieved by using a two-steps procedure: the registration of pairs of range images between them, and the posterior minimization of the global error for the whole set of images.

Date of Award	28 Nov 2013
Original language	English
Supervisor	Xavier Binefa Valls (Director)