| Computer vision involves inferring the state of the world from noisy and inherently ambiguous images. Some of the major challenges include object recognition: finding a known object in an image; segmentation; separating an image into different regions; tracking; recognizing and extracting desired objects as they move spatially in a sequence of images. The techniques developed to perform the aforementioned tasks are used in many applications like face recognition, extracting out a specific structure (like a corpus callosum or ventricles) from the MRI/CT images of a human brain and following an object (like a vehicle or a man) in a video sequence. Extensive research is being done to come up with automatic algorithms that can segment or track or recognize an object. For a human observer, these tasks seem easy because of the use of prior knowledge about the task at hand. More specifically, prior knowledge about the shape of an object can be quite useful in all the above applications.; This thesis deals with the problem of tracking highly deformable objects in the presence of noise, clutter and occlusions. The contributions of this thesis are threefold: (1) A novel technique is proposed to perform filtering on an infinite dimensional space of curves for the purpose of tracking deforming objects. The algorithm combines the advantages of particle filter and geometric active contours to track deformable objects in the presence of noise and clutter. (2) Shape information is quite useful in tracking deformable objects, especially if the objects under consideration get partially occluded. A nonlinear technique to perform shape analysis, called kernelized locally linear embedding, is proposed. Furthermore, a new algebraic method is proposed to compute the pre-image of the projection in the context of kernel PCA. This is further utilized in a parametric method to perform segmentation of medical images in the kernel PCA basis. (3) The above mentioned shape learning methods are then incorporated into a generalized tracking algorithm to provide dynamic shape prior for tracking highly deformable objects. The tracker can also model image information like intensity moments or the output of a feature detector and can handle vector-valued (color) images. |
