Efficient and effective automated surveillance agents using kernel tricks

Abstract

Many schemes have been presented over the years to develop automated visual surveillance systems. However, these schemes typically need custom equipment, or involve significant complexity and storage requirements. In this paper we present three software-based agents built using kernel machines to perform automated, real-time intruder detection in surveillance systems. Kernel machines provide a powerful data mining technique that may be used for pattern matching in the presence of complex data. They work by first mapping the raw input data onto a (often much) higher-dimensional feature space, and then clustering in the feature space instead. The reasoning is that mapping onto the (higher-dimensional) feature space enables the comparison of additional, higher-order correlations in determining patterns between the raw data points. The agents proposed here have been built using algorithms that are adaptive, portable, do not require any expensive or sophisticated components, and are lightweight and efficient having run times of the order of hundredths of a second. Through application to real image streams from a simple, run-of-the-mill closed-circuit television surveillance system, and direct quantitative performance comparison with some existing schemes, we show that it is possible to easily obtain high detection accuracy with low computational and storage complexities.