Source:
Department of Electrical Engineering and Computer Science, Northwestern University, Evanston, p.85 (2006)
Abstract:
Two fundamental tasks in computer vision and video processing applications are detection
of foreground objects by subtracting the background and video segmentation. In this thesis we
first present an overview of the approaches in the literature that are proposed to solve these
challenging problems. This is followed by presentation of two novel spatiotemporal algorithms
applicable to these problems with a stationary single camera systems. In the first part, we
present a novel probabilistic background modeling and subtraction method that exploits spatial
dependencies between pixels. By using an initial segmentation of the background scene, we
model each pixel by a mixture of spatiotemporal Gaussian distributions.
Using the ideas of the background model, we present a novel algorithm for segmenting
video sequences into objects with smooth surfaces. The segmentation of image planes in the
video is modeled as a spatial Gibbs-Markov random field and extension to 3D is accomplished
by the temporal labeling algorithm. Experimental results for indoor and outdoor surveillance
videos demonstrate the performance advantages of the proposed methods. Possible extensions
and applications are outlined and conclusions are drawn.
