Statistical and geometric methods for shape-driven segmentation and tracking
MetadataShow full item record
Computer Vision aims at developing techniques to extract and exploit information from images. The successful applications of computer vision approaches are multiple and have benefited diverse fields such as manufacturing, medicine or defense. Some of the most challenging tasks performed by computer vision systems are arguably segmentation and tracking. Segmentation can be defined as the partitioning of an image into homogeneous or meaningful regions. Tracking also aims at extracting meaning or information from images, however, it is a dynamic task that operates on temporal (video) sequences. Active contours have been proven to be quite valuable at performing the two aforementioned tasks. The active contours framework is an example of variational approaches, in which a problem is compactly (and elegantly) described and solved in terms of energy functionals. The objective of the proposed research is to develop statistical and shape-based tools inspired from or completing the geometric active contours methodology. These tools are designed to perform segmentation and tracking. The approaches developed in the thesis make an extensive use of partial differential equations and differential geometry to address the problems at hand. Most of the proposed approaches are cast into a variational framework. The contributions of the thesis can be summarized as follows: 1. An algorithm is presented that allows one to robustly track the position and the shape of a deformable object. 2. A variational segmentation algorithm is proposed that adopts a shape-driven point of view. 3. Diverse frameworks are introduced for including prior knowledge on shapes in the geometric active contour framework. 4. A framework is proposed that combines statistical information extracted from images with shape information learned a priori from examples 5. A technique is developed to jointly segment a 3D object of arbitrary shape in a 2D image and estimate its 3D pose with respect to a referential attached to a unique calibrated camera. 6. A methodology for the non-deterministic evolution of curves is presented, based on the theory of interacting particles systems.