Fast active object tracking in color video

An important problem in image analysis is object segmentation and tracking in image sequences. It involves the isolation of a single object from the rest of the image that may include other objects and a background. Here, we focus on boundary detection of one or several objects by a dynamic model known as the `geodesic active contour´. Although the geodesic active contour model has many advantages over its predecessors-snake model and geometric flow, its main drawback is its non-linearity that results in inefficient implementations. For example, explicit Euler schemes for the geodesic active contour limit the numerical step for stability. In this paper we introduce a new method that maintains the numerical consistency and makes the geodesic active contour model computationally efficient. The efficiency is achieved by cancelling the limitation on the time step in the numerical scheme, by limiting the computations to a narrow band around the the active contour, and by applying an efficient re-initialization technique