- 作者: Jinshiuh Taur
- 中文摘要: Three-dimension shape reconstruction is one of the important research areas in object recognition and image understanding. A structure-from-motion problem as originally proposed by C. Tomasi and T. Kanade in 1992 has attracted a lot of attention. It is based on the singular value decomposition (SVD) approach. In this paper, it is extended to cope with the multi-target case. That is, given a sequence of 2-D video images of multiple moving targets, the goal is to compute the 3-D motion of the targets and reconstruct their 3-D shapes. This is further extended to the multi-camera-multi-target problem. First, a robust algorithm which enhances the reliability of the block matching techniques is proposed for fast tracking of feature points in a sequence of images. Then the feature points are mapped onto their corresponding objects using an algebraic method based on the subspace clustering method and principal singular vector (PSV). Thereafter, the motion and shape may be estimated from a matrix factorization using SVD. We demonstrate the effectiveness of the algorithms in tracking and reconstruction of the shape information using both artificially created data and a real image sequence in somewhat controlled environments.
- 英文摘要: --
- 中文關鍵字: singular value decomposition, feature tracking, shape reconstruction from motion, image understanding
- 英文關鍵字: --