Efficient 3D visual perception for intelligent surveillance and autonomous navigation지능형영상감시와 자율주행을 위한 효율적인 3차원 영상인지기법

Cited 0 time in webofscience Cited 0 time in scopus
  • Hit : 801
  • Download : 0
An image is a 2D projection of the 3D world. Therefore, it is not straightforward to restore 3D information from the 2D image. This thesis is based on three observations available in our urban environments. First, an object is usually on a plane such as a floor and a road. Second, a surveillance camera and dashboard camera attached to a car and robot have a constant height from the ground. Third, Two adjacent images captured by a camera has little change. This thesis establishes three major constraints based on the observations and derives two efficient geometric tools. From a single image, on-plane projective geometry is a tool to interpret metric properties of an object on a plane. It is based on a canonical camera which is a virtual camera aligned to the plane. The canonical camera makes complex 3D inference problems simple and accessible. The on-plane projective geometry was applied to metric image processing and metric intelligent visual surveillance. From an image sequence, simplified epipolar geometry provides methods to estimate relative pose between two images based on motion constraints. It includes fast and reliable 2-point relative pose solvers for planar motion. It also contains fast and iterative 5-point relative pose solver for general but small inter-frame motion. The relative pose solvers were applied to monocular visual odometry and demonstrated their efficiency. Monocular visual odometry and SLAM suffer from motion scale ambiguity, which is tackled by two approaches. First, the motion scale is retrieved from the camera's constant height constraint. Second, the motion scale is measured by an additional sensor, speedometer. This thesis demonstrates scale-corrected monocular visual odometry whose performance is close to the state-of-the-art.
Advisors
Kim, Jong Hwanresearcher김종환researcher
Description
한국과학기술원 :로봇공학학제전공,
Publisher
한국과학기술원
Issue Date
2019
Identifier
325007
Language
eng
Description

학위논문(박사) - 한국과학기술원 : 로봇공학학제전공, 2019.2,[vi, 71 p. :]

Keywords

3D vision▼aprojective geometry▼aepipolar geometry▼amulti-view geometry▼acamera pose estimation▼aintelligent visual surveillance▼avisual odometry; 3차원비전▼a투영 기하학▼a에피폴라 기하학▼a다시점 기하학▼a카메라 자세 추정▼a지능형 영상 감시▼a영상 오도메트리

URI
http://hdl.handle.net/10203/264614
Link
http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=842005&flag=dissertation
Appears in Collection
RE-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