고밀도의 정확한 3차원 포인트 클라우드 획득 시스템에 관한 연구
A Study on Accurate and High Density 3D Point Clouds Capture System
- 주제(키워드) 3D Point , Sturctured Light , Registration , Accurate , Density
- 발행기관 서강대학교 영상대학원
- 지도교수 서용덕
- 발행년도 2017
- 학위수여년월 2017. 2
- 학위명 박사
- 학과 및 전공 영상대학원 미디어공학
- 실제URI http://www.dcollection.net/handler/sogang/000000061350
- 본문언어 영어
- 저작권 서강대학교 논문은 저작권보호를 받습니다.
초록/요약
We present a framework for multi-camera and multi-projector object acquisition based on structured light that can measure millions of 3D point clouds and the reflection models of real objects rapidly and stably. Active techniques for obtaining the shape of an object, such as structured light and time-of-flight (TOF) lasers, provide a robust and sophisticated three-dimensional shape. Such methods have the advantage that the reliability of the correspondence is guaranteed. The structured light can obtain appropriate geometric information for an object, but its detailed shape tends to be averaged or its horizontal resolution is reduced by the limited resolution of a projector. To overcome this limitation, various methods for estimating the subpixel of the projector have been proposed in recent years. On the other hand, the pose estimation is required for shapes obtained from several different viewpoints in order to create a 3d model of the object. An iterative close point (ICP) algorithm, a typical method for registering 3D point cloud, estimates the rigid transformation between points at two different viewpoints using a least square method. But this method requires a good initial solution for a convergence and the solution is only locally searched in the 6 dimensional space, which result in unsatisfactory registration in many cases. Recently researches on global optimization have been reported, but its computational time complexity is too high to get a solution in a reasonable amount of waiting time. Here, we propose two methods to improve the accuracy of the acquisition system of the 3D point clouds and to register 3D point clouds rapidly. The first proposal is the super resolution method using multiple cameras and multiple projectors to improve the accuracy of shape recovery. Through the statistical analysis, we show a novel super-resolution scheme which utilizes projectors in different positions. The structured pattern at different viewpoints is projected and superimposed so that a minimal area on the surface can be uniquely identified. Therefore, it is possible to obtain a high dense and accurate 3D point clouds. Then, we can estimate the position of the projectors by using the acquired 3D point clouds and also measure the reflection property of the object. The second one is to physically reduce the search space by using a calibration process designed by a turntable and mathematical analysis, in order to consider not only the registration performance but also the computational complexity of building the 3D model of the object. This method proves that there is a geometric transformation between structured light and a turntable, and shows that the registration of transformed point clouds can be performed rapidly through one-dimensional rotation space search.
more