A Foreground Detection System for Automatic Surveillance

Automated surveillance has long been an application goal of computer vision. An integral part of such surveillance systems is concerned with accurately segmenting foreground objects from the static background in the videos. In this thesis we introduce a novel system for background subtraction, which takes a different approach than the conventional background subtraction systems. We make the assumption that the video background is stationary and the foreground objects take up only a small portion of the entire frame at any given time. This specific assumption allows us to formulate the foreground signal as a sparse additive error introduced to otherwise clean background signal. We outline the algorithm for performing background subtraction using linear programming, and demonstrate accurate segmentations of foreground objects under realistic surveillance scenarios. The proposed method is on par with the state of the art approaches for accurately segmenting the foreground under challenging conditions. Furthermore we propose several methods for building a set of bases to represent the background and provide empirical justification of their effectiveness.