Tracking and Describing Deformable Objects Using Active Contour Models

Technical Report TR-CIM-90-9 In this thesis we consider a number of issues in developing techniques and algorithms to automate the visual tracking of deformable objects in the plane. We have applied these techniques in cell locomotion and tracking studies. We examine two classes of computer vision problems. First, we consider the segmentation of a noisy intensity image and the tracking of a nonrigid object. Second, we consider the shape analysis of an amorphous object. In evaluating these problems, we explore a new technique based on an active contour model commonly called a "snake." The snake permits us to simultaneously solve, in constrained cases, both the segmentation and tracking problems. We present a detailed analysis of the snake model, emphasizing its limitations and shortcomings, and propose various improvements to the original description of the model. Then, we study the two complementary types of shape descriptors: boundary- and region-based. We propose to combine these within the context of the grassfire transform. Two new algorithms are described. First, we present a contour segmentation technique using mathematical morphology on the curvature function, we call curvature morphology. Accurate localization for different scales of curvature features is achieved, leading to a Morphological Curvature Scale-space or MCS. Second, the snake model is used to simulate the grassfire transform using the previously extracted contour features. This permits us to produce a multiscale skeleton representation of shape which is based on the Euclidean distance metric. New significance criteria for our shape descriptors, such as the ``region-support'' of curvature extrema and the ``ridge-support'' of skeleton branches are also proposed. Finally, numerous implementation details are discussed; for example, the description of an efficient sequential Euclidean distance transform