This thesis is focused on visual tracking by density matching flows and its generalization to 3D medical image segmentation.; Visual tracking is the problem of estimating the motion or positions of an object given a sequence of images. It has extensive applications in autonomous robotics, surveillance and image analysis. The major challenges in visual tracking include how to track a non-rigid object in a cluttered or dynamic background. A novel tracking method based on density matching is proposed to deal with above challenges. Unlike most existing methods, the method does not employ edges as features, and does not assume Gaussian observation probability. Instead, the method aims at tracking a non-rigid object moving in clutter using photometric information. In this method, the object is represented as a set of curves; the prior knowledge about the object is represented as a model density of photometric variables. In the tracking process, the curves move in directions minimizing the distance between a sample density and the model density. The distance may be the Kullback-Leibler information number or the Bhattacharyya measure. Depending on how the sample density is selected, three variants of the method can be derived. A shape prior term can be incorporated to improve robustness of the method. These methods are formulated using partial differential equations (PDEs) and are solved numerically by level sets. Comparison experiments show the method tracks well.; Medical image segmentation is one of the most heavily investigated fields due to its potential applications. Most CT images have low quality, and have no sharp edges. Our tracking method is generalized to a 3D segmentation method for medical image segmentation. In the new method, besides a model distribution, shape priors are represented by a point-based principle component analysis (PCA) model. The shape model may be a multiple object model. The model distribution and the shape model are coupled through a group of variables. By minimizing the distance between the model distribution and an empirical distribution (computed during the segmentation process) in terms of the group of variables, an ordinary differential equation (ODE) formula is derived. Our segmentation method employs distribution matching. Distribution matching does not require the difficult, time-consuming pixelwise computation between the model and the image. The method is applied to prostate and rectum segmentation. The results are promising.
展开▼
