Incremental Svdd Training: Improving Efficiency of Background Modeling in Videos

Tracking moving objects in videos with quasi-stationary backgrounds is one of the most important and challenging tasks in video processing applications. In order to detect moving foreground regions in such videos the background and its changes should be modeled to help detecting moving regions of interest. Support Vector Data Descriptors (SVDD) can be employed in order to analytically model the background and explicitly account for its inherent changes. The major draw back of the SVDD modeling is the issue of training of the SVDD which is a quadratic programming (QP) problem. In this paper we propose a method to efficiently train the SVDD’s. The advantages of our technique are its low memory requirement and its efficiency in terms of speed. The proposed method runs in constant time with respect to the size of the training data set since its retraining is performed only on the support vector working set.