Single-class SVM for dynamic scene modeling

Scene modeling is the starting point and thus the most crucial stage for many vision-based systems involving tracking or recognition. Most of the existing approaches attempt at solving this problem by making some simplifying assumptions such as that of a stationary background. However, this might not always be the case, as swaying trees or ripples in the water often violate these assumptions. In this paper, we present a novel method for modeling background of a dynamic scene, i.e., scenes that contain “non-stationary” background motions, such as periodic motions (e.g., pendulums or escalators) or dynamic textures (e.g., water fountain in the background, swaying trees, or water ripples, etc.). The paper proposes single-class support vector machine (SVM), and we show why it is preferable to other scene modeling techniques currently in use for this particular problem. Using a rectangular region around a pixel, spatial and appearancebased features are extracted from limited amount of training data, used for learning the SVMs. These features are unique, easy to compute and immune to rotation, and changes in scale and illumination. We experiment on a diverse set of dynamic scenes and present both qualitative and quantitative results, indicating the practicality and the effectiveness of the proposed method.