Modeling Local Video Statistics for Anomaly Detection
Roman Garnett · Open Scholarship Institutional Repository (Washington University in St. Louis) · 2004
This paper promotes a probabilistic approach for building models of local video statistics for use in background subtraction schemes. By shifting into a probabilistic framework, additional analytical tools become available for the creation and evaluation of these models. This paper continues to suggest the use of nonparametric statistical methods for measuring the quality of efficient local spatio-temporal models of video background distributions. Beginning with the familiar relative entropy distance between probability distributions, we create a new distance measure that can be used to quantitatively measure the quality of a probabilistic background model.