Motion Segmentation And Motion Estimation Based On2D/3D Videos

Motion segmentation and motion estimation are two basic problems of image processing and computer vision, which are widely used in many areas. This thesis focuses on motion analysis based on2D/3D videos, and proposes a variational model combined with active contour evolution and motion parameter estimation, hoping to find robust solutions for motion segmentation and motion estimation with moving viewing system and multiple objects. In summary, the major work and main contributions of this thesis are listed as follows:●We propose a unified framework for motion segmentation and motion estimation with moving viewing system and multiple objects based on2D/3D videos. First, a spatiotemporal energy functional is built up to perform motion segmentation and estimation simultaneously. The objective function is integrated throughout the spatiotemporal domain, thus the trajectory of moving target’s boundary forms a surface in the spatiotemporal domain, which also achieves the tracking of moving objects. Moreover, the combination of motion segmentation, tracking and parameter estimation can effectively use the spatial and temporal information for better three-dimensional motion constraints. Second, it allows segmentation with a moving viewing system and does not need prior information of camera motion; In addition, the proposed method has no limitation with respect to the number of targets, which means any object that has non-homogeneous motion with the background can be partitioned and tracked.●We propose a new model for the spatiotemporal segmentation model based on total variation, to overcome the low computational efficiency and sensitive initialization problem with the level set method, and prove the intimate connection between the spatiotemporal segmentation models based on active contour and total variation. We introduce a convex optimization technique to convert the original energy functional into a global convex functional, and get rid of the "curse of non-convexity".The convexity of segmentation variable provides the possibility that the segmentation results are no longer dependent on initial conditions, and any optimization method can guarantee the global optimal solution. Compared to discrete algorithms such as Graph-Cuts, this continuous model provides faster and more accurate solutions.●The numerical bottleneck of the convex-optimized spatiotemporal segmentation model comes from the total variation part in energy functional that suffers from serious nonlinearity and non-differentiability. To overcome the shortcomings and deficiencies of the traditional method for solving the variational model, we propose two fast and efficient algorithms--primal-dual algorithm and Split-Bregman algorithm.●We propose a new convex model for motion segmentation and dense three-dimensional interpretation based on monocular image sequences. First, a constraint for motion parameters and depth information is proposed and we use use Bayes method to model the energy functional. Second, a convex relaxation method is applied to the proposed model, overcoming the limitation of the local minimum problem during functional minimization. At last, an alternate iterative procedure is adopted to solve the multi-variable optimization problem, respectively estimating motion parameters, estimating depth and evolving surface after initialization.