High-quality 3D reconstruction by cameras카메라를 이용한 고품질 3차원 복원 방법

Cited 0 time in webofscience Cited 0 time in scopus
  • Hit : 652
  • Download : 0
3D reconstruction is a computer vision technique to restore 3D structures of scenes from digital images. The three factors that determine the quality of reconstruction are camera model, camera pose, and correspondence. For high-quality 3D reconstruction, it is paramount important to know these factors accurately. We categorize various techniques in 3D reconstruction according to which factors are mainly handled and notice that the techniques in different categories can be combined to achieve complete 3D reconstruction in three different ways. In this dissertation, we explore the three ways of 3D reconstruction and introduce new solutions to their critical applications including 3D scanning, see-through car, and depth from small motion. Furthermore, we develop a new method for camera calibration that uses a display device to unleash the camera calibration technique from the restraint of using blurred images by camera defocus. Firstly, we present our camera calibration method, which is one of the most significant contributions of this thesis. Unlike traditional approaches for camera calibration that necessarily require well-focused images, our method can perform accurate calibration with heavily defocused images. With a set of binary patterns that we propose, we define a feature point as an intersection of a vertical edge and a horizontal edge. We then formulate the feature extraction as a deconvolution problem with a sub-pixel level feature location and a Gaussian blur kernel. Furthermore, we compensate for camera parameter error due to refraction of the glass panel of a display device. We evaluate the performance of the proposed method on both synthetic data and real data. Secondly, we present two approaches that are based on structured light: a single-shot based 3D scanning, and an accurate multi-view matching for highly accurate 3D scanning, respectively. In the single-shot method, we recover not only the 3D structures but also the intrinsic properties of a scene via an efficient parameterization of the image model, based on a color-coded phase-shifting pattern. In the multi-view method, we try to estimate the correspondence between multi-camera images as accurate as possible so to achieve very precise 3D reconstruction. Our method estimates the optimal depth for each pixel of each camera which makes the difference between the decoded pattern values at its projections across multiple views be minimized. We build a multi-camera 3D scanning system to demonstrate the superior performance of the proposed method. Thirdly, we present the first real-time system to see through a car seamlessly. When two vehicles are equipped with an appropriate wireless communication system, the stereo vision system mounted on the front car allows creating a sparse 3D map of the environment where the rear car can be localized. Based on the pose between the cars, an augmented-reality image is synthesized and transferred to the rear car, which can replace the occluded field of rear car view and create a seamless see-through effect to help the rear car driver to prevent a potential hazard and overtake the front car safely. Finally, we present our methods to recover a depth map from a small motion video clip. We first study interesting characteristics of small motion assumption in the case of a calibrated camera, and then present a method for high-quality depth estimation for uncalibrated cameras, like our daily mobile phones. As opposed to prior methods that only recover scene geometry and camera motions, we introduce a self-calibrating bundle adjustment tailored for small motion. For dense matching, our variance-based cost function has been utilized to leverage the benefit of having negligible intensity changes within the scene due to the minuscule variation in viewpoint. The depth maps obtained by the proposed framework show extremely fine structures that are unmatched by previous literature under the same small motion configuration.
Advisors
Kweon, In Soresearcher권인소researcher
Description
한국과학기술원 :전기및전자공학부,
Publisher
한국과학기술원
Issue Date
2017
Identifier
325007
Language
eng
Description

학위논문(박사) - 한국과학기술원 : 전기및전자공학부, 2017.8,[viii, 128 p. :]

Keywords

3D reconstruction▼a3D scanning▼acamera calibration▼astructured-light▼asee-through car▼adepth from small motion; 3차원 복원▼a3차원 스캐닝▼a카메라 캘리브레이션▼a구조광▼a차량 투시▼a작은 움직임

URI
http://hdl.handle.net/10203/242060
Link
http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=718944&flag=dissertation
Appears in Collection
EE-Theses_Ph.D.(박사논문)
Files in This Item
There are no files associated with this item.

qr_code

  • mendeley

    citeulike


rss_1.0 rss_2.0 atom_1.0