SVDD-Based Video Anomaly Detection Approaches

In this chapter, we introduce a support vector data description (SVDD)-based method for anomaly detection in video. Support vector data description is a single-class nonparametric method that finds the boundary around a dataset that encloses the normal samples so as to establish the separation between two classes, that is, the normal events and the abnormal events in the video anomaly detection applications, that is, models the support of the distribution. The SVDD has a number of benefits including that it can model arbitrarily-shaped multi-modal distributions (namely, a non-Gaussian modeling), fewer training samples are needed to characterize the background scene in high-dimensional spaces (sparsity), and that the method avoids overfitting and yields good generalized results (i.e., good generalization ability). As an extension of one-class support vector machine (SVM), for anomaly detection, it estimates a hypersphere that contains the most normal data in the input space, outside of which the abnormal data should reside. It works by find a kernel function that maps samples into a higher dimensional space to obtain a better discrimination.