Smart Cameras for Visual Surveillance

Abstract

Automated visual surveillance systems must retrieve and integrate relevant information from a large number of sensors and present it in a user-friendly manner. These systems rely on smart cameras that extract target meta-information from content-rich video data by detecting and classifying the targets of interest in their respective fields of view and tracking them over time. The smart cameras thus provide an ideal platform for data collection and low/mid-level video analysis. However, high-level scene understanding and context-sensitive reasoning require learning of environment parameters and their spatiotemporal variation. Such learning and reasoning is an important characteristic ability of all cognitive systems that increase the adaptability and thus the utility of such systems. In this chapter, we show how smart cameras can mimic this intelligent behavior by learning models from past observations and exploiting them to enable robust target detection and classification, persistent target tracking, anomalous behavior detection, inter-sensor calibration, and geo-registration.