| Description |
Image segmentation and object tracking are two of the most fundamental computer vision problems. In both segmentation and tracking, how we represent or model the shape of the object of interest has significant effect on the performance of these algorithms. Existing shape models are generally either explicit-parametric (e.g. active shape models), or they are implicit-nonparametric (e.g. level set methods). Implicit-parametric shape models are rarely studied. In this dissertation, we present a Disjunctive Normal Shape Model (DNSM) which is a differentiable implicit and parametric model. We focus on formulations of the DNSM in a Bayesian inference framework to solve some of the major limitations of the existing image segmentation and object tracking methods. First, we develop a disjunctive normal appearance and shape model for image segmentation. The use of priors in image segmentation is known to improve accuracy; however, existing techniques either require landmark points, or they do not handle topological changes, or they do not support construction of local priors. In our framework, the DNSM's parametric nature allows the use of powerful local prior statistics, and its implicit nature removes the need to use landmarks and easily handles topological changes. Second, we present a novel disjunctive normal parametric level set method for both two-phase and multiphase image segmentations. Level set methods are widely used for image segmentation because of their capability of handling topological changes. Compared to the conventional level sets, the proposed method has significantly less computational cost, it naturally keeps the level set function regular, it is more suitable for multiphase and local region-based image segmentations, and it is less sensitive to noise and initialization. Finally, we develop a DNSM-based online object tracking method. Most object tracking algorithms use single or multiple bounding boxes to represent and track objects. However, this simple target representation results in inclusion of background noise in the object appearance modeling, and they usually do not allow the use of shape information. We propose a robust object tracking algorithm by integrating the DNSM, for shape priors, with the Kernelized Correlation Filters (KCF), for appearance modeling, in a Bayesian inference framework. |
