Content Analysis of Distributed Video Surveillance Data for Retrieval and Knowledge Discovery

Acknowledgements My thanks go to all colleagues, friends and my family, especially to 1 Introduction A video surveillance activity has dramatically increased over the past five years [Tri05]. It was caused mainly by the escalating amount of surveillance cameras as a request to the terrorist peril. However, all of the image understanding and risk detection is left to human. An automated system for visual event detection and indexing can reduce the burden of continuous concentration on monitoring and increases the effectiveness of information reuse by the security, police, emergency and firemen. Additionally – in contrast to the (relatively) cheap technical equipment, the work of security personnel is very expensive in developed countries. In the last decade also the machine vision research has (in some aspects) emerged to practical applications running on present computers in real time. There are special applications that can track [CLEAR] and count people in single camera's field of view [Axis] or detect a left luggage [PETS06]. Some other commercially successful applications deal for instance with traffic [Camea], face [COGNI] or other biometrics [L1id] detection, recognition and identification. The output of such applications is a (textual) features' and semantics' annotation in a form of a relational [VACE07] or XML data [MP704], potentially connected to an alarm. An example of such process including some other (prior) knowledge is in Fig. 1.1. 3 P e and the analysis of behavior for an automated detection of special events including (false) fire alarms. Additionally, an application of data mining in databases containing information about objects and their spatio-temporal location might prevent such events.