Motion estimation of human from a single or a monocular sequence of image is an important research field in Computer Vision and Image Processing. In this paper, the following topics are investigated: (1)Based on coplanar constraint and rigidity constraint, we proposed a novel approach for 3D-motion analysis of coplanar multi-links based on feature points correspondences. We got the conclusion that 3D motion and structure of the coplanar jointed links can be determined from a monocular sequence of image, and the number of frames needed is related with not only the number of rigid links, but also the relative position of the space unknown plane. Furthermore, we applied this technique to the human 3D-motion estimation. To achieve the robust 3D-motion, we introduced the Genetic Algorithm. (2) We proposed a novel 3D-motion estimation method for one kind of special multi-link -- the product of exponential maps based on region correspondences. By integrating the special motion model of the exponential links into the region-based 3D-motion estimation, we yielded a general method for the 3D-motion estimation of exponential link model. The advantages of this particular formulation are that it results in the equations that needed to be solved for frames being linear and the solutions to equation are the closed form. The results are more robust than that based on feature point correspondences. The technique was validly applied to 3D-motion estimation of fingers. (3) Based on constraint fusion: coplanar constraint, rigidity constraint, smooth motion constraint, we proposed the 3D motion estimation from the monocular perspective images without any special markers. We considered the arm as stick model, and by pre-processing the image sequence wesemi-automatically yielded 2D correspondences of joint points of arm among images. Then by constraint fusion 3D relative structure of arm can be determined in a scale factor. The experiment with real images is included.
修改评论