Ph.D. Theses

Visual Tracking by Density Matching: Theory, Algorithms and Application

By Tao Zhang
Advisor: Daniel Freedman
May 6, 2005

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 robots, surveillance, and medical image analysis. The major challenges in visual tracking include how to track a non-rigid object in cluttered backgrounds or dynamic backgrounds. A novel tracking method based on density matching was proposed to deal with above challenges. Unlike most existing methods, the method does not employ edges as features, does not assume Gaussian observation probability. Instead, the method aims at tracking a non-rigid object moving in cluttered backgrounds using photometric information (such as color, texture). In this method, the object is represented as 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 Kullback-Leibler information number or Bhattacharyya measure. Depending on how to select the sample density, three variants of the method can be derived. A shape prior item can also be incorporated to improve the robustness of the method. These methods are formulated by Partial Differential Equations (PDEs) and are solved numerically by level sets. Comparison experiments show the method has strong ability in tracking.

In the second part, we developed a 3D medical image segmentation method. Medical image segmentation is one of the most heavily investigated fields due to its potential applications. Most medical images are in low quality, have no sharp edges. In this thesis, our tracking method was further generalized to a 3D segmentation method for medical image segmentation. In the 3D segmentation method, besides a prior model density, shape priors are represented by a point-based PCA model. The model density and the shape are coupled through a group of variables. By minimizing the model density and a sample density (sampled from a given shape) in terms of the group of variables, an Ordinary Differential Equation formula is derived. The method was applied to prostate and rectum segmentation. The results were promising.

Return to main PhD Theses page