One-Class Background Model Assaf Glazer, Michael Lindenbaum, Shaul Markovitch {assafgr,mic,shaulm}@cs.technion.ac.il Technion – Israel Institute of Technology, Haifa, Israel Abstract. Background models are often used in video surveillance sys- tems to find moving objects in an image sequence from a static camera. These models are often built under the assumption that the foreground objects are not known in advance. This assumption has led us to model background using one-class SVM classifiers. Our model belongs to a fam- ily of block-based nonparametric models that can be used effectively for highly complex scenes of various background distributions with almost the same configuration parameters for all examined videos. Experimen- tal results are reported on a variety of test videos from the Background Models Challenge (BMC) competition. 1 Introduction Moving foreground objects in an image sequence from a static camera can be detected by comparing new images with a representation of the background scene. This process is called background subtraction and the representation of the background is called the background model. Background models must be able to cope with changes in the background scene that may occur over time. These include illumination changes, fluctuations of local image patterns (e.g., swaying trees and fluttering flags), flickering CRTs, and so on. A common assumption in background modeling is that ground-truth images are not available for training. Thus, background models should be built with- out knowledge about what foreground objects are expected to appear [1]. With respect to the background, however, it is common to assume that a priori knowl- edge is available for training. Because the available information pertains to only one side of the problem, we propose in the following to use one-class classification tools to model background scenes. In general, traditional supervised classifiers are trained using positive and negative examples. However, in our settings, labeled data exist for only the background class. A straightforward learning approach would be to estimate the distribution of the background. However, density estimation in high-dimensional data is hard, requiring a large number of examples, and is sensitive to outliers. One-class classifiers are an efficient alternative. Unlike the binary decisions out- put by traditional classifiers, the decisions output by one-class classifiers tell us whether examples were drawn from the distribution of the learned class. In this work we use one-class SVM (OCSVM ) classifiers [2] to model the distribution of the background. Our decision to use OCSVM is motivated by its nonparametric