|
Multi-view range image registration is a technique of bringing range images obtained from different camera coordinate systems to a common coordinate system. It is one of the important steps to generate complete 3D models of real objects from 3D range sensors. In recent years, many investigations have addressed multi-view range image registration. However, it is usually a time-expensive and difficult task whose accuracy affects the quality of final 3D models. Thus, conventional investigations usually focus on accurate implementation of registration refinement. Registration refinement is also a 3D shape matching technique which finds correspondence between different 3D shapes. It usually needs considerable computation because that 3D data from a range sensor consists of hundreds of thousand of points. In this reason, few systems have addressed real-time registration problem. In this paper, we propose a real-time and on-line 3D registration system which acquires and registers multi-view range images simultaneously. Continuously obtained range images from a hand-held range sensor are registered by using a geometric refinement techniques run on a GPU. To register range images in real-time, we employ a GPU(Graphic Processing Unit) programming technique. GPU is a processor for graphic work originally, however recent computer vision researches employ GPU for fast and real-time implementation of computer vision algorithms such as KLT tracker and SIFT. We implement a point-to-plane registration refinement technique using GPU. To use up-to-date techniques of GPU, we use the CUDA architecture which is available beyond Geforce 8 series graphic boards from NVIDIA. Most linear algebra used in the refinement process is implemented by GPU programming except a least squares minimization. For experiments with real objects, a hand-held stereo camera is used to
continuously obtain range images. Results in this extended abstract show the proposed system can registered the range images in very fast time.
|
