An Optical Flow based Method for the Segmentation of Image Sequences

Abstract

The segmentation of motion in an image sequence is an important task in many computer vision applications. This thesis presents the theory and numerical algorithm for the detection of object boundaries of moving objects in an image sequence, using optical flow to estimate movement and active contours to locate flow boundaries. Three different methods for regularizing optical flow are presented and analyzed based on the application of this flow field in a segmentation framework. This active contours framework is formulated as the computation of geodesic curves in a Riemannian space defined by the gradients of the optical flow field. We will use a level set function to describe and evolve these contours, effectively incorporating structural information of the flow field in the level sets of this function. A Mumford-Shah segmentation is used to extract this information from the flow field. The boundary of this segmentation is represented by a cubic B-spline. This leads to a combined evolution of the level set function and the B-spline curve. The performance of the algorithm is validated on two real-world data sets, provided by the Norwegian Defence Research Establishment (FFI).