The Research Of Non-rigid Object Motion Reconstruction Based On Dynamic Image Sequences

One of the important study of computer vision is recovering the 3D structure and motion from dynamic image sequences. The factorization method based on dynamic image sequences has the good robustness and accuracy, so the method is widely use to solve the problem of 3D motion reconstruction. The traditional factorization methods work mainly for static rigid object, it has difficult using the methods directly to solve the problem of structure and motion of non-rigid object under dynamic sense.At present, the methods of non-rigid 3D motion reconstruction all assume that the number of the non-rigid object shape base is known. However, the number of the shape base is very important to the method of 3D reconstruction. If the number is failure, the 3D reconstruction method will be invalid. Otherwise previous methods are all based on the assumption of affine camera model. This is an approximation to the real perspective projection model, and this assumption is only valid when the object is small compared to the distance between the object and the camera. The assumption will bring large reconstruction errors when the object is close to the camera.Aimed at above problems, the main work of this paper are following:(1) The paper proposes the estimation method of the deformation degree of the non-rigid object motion from image sequence. Estimation the deformation degree is very important to non-rigid structure-from-motion from image sequence. The existing deformation degree estimation methods usually can not deal with the missing image point problem and assume the feature points have no position noise or the uncertainty is isotropic. As a result, these methods generalize badly. To improve generalization, we provide a method for estimating deformation degree by taking into account the factor of directionality uncertainty and missing data. Experimental results from human activities image sequence demonstrate the feasibility of the proposed approach.(2) This paper studies a improved linear iterative algorithm to solve the problem of recovring the non-rigid 3D structure and motion information from dynamic image sequences under the real perspective projection model. Existing algorithms all assume the camera model is a weak perspective projection model, the assumption is only valid when the size and depth change of the object is very small compared to the distance between the object and the camera. So the paper extendes factorization method from the weak perspective projection model to the general perspective projection model using the linear iterative algorithm. Experiments to the dynamic image sequences of facial expression change shows that the improved algorithm is the accuracy and effectiveness.(3) Apply the power factorization method to solve the problem of non-rigid 3D motion reconstruction based on dynamic image sequences. Factorization method based on singular value is more complex for solving the transformation matrix, and there are large reconstruction errors in the method. So the paper studys the power factorization to solve the problem of recovering non-rigid object 3D structure and motion information from the dynamic image sequences. Experiments on face image sequences show that the reconstruction result of the power factorization algorithm is more accurate than the result of the factorization based on singular value.