On-Line Object Behaviour Analysis for Surveillance Systems

Automatic surveillance systems are an important
emerging application of on-line simple or compound event
detection algorithms in video or audio data. The nature of
such systems implies several requirements on the used
algorithms. The system ability to give the response on-line
is the main topic addressed in this work. The paper
predefines the requirements for each system recognition
module that must work in real-time or faster.
This paper describes and analyses several event
recognition algorithms based on video processing,
concerning the applicability in on-line surveillance
systems. Two main trends are addressed with focus on the
methods' speed and robustness. Statistical approaches are
the basis for simple event detectors whereas spatiotemporal
rules define compound event detectors.
Examples of statistical approaches are ad-hoc bicycle
detector, dog detector based on AdaBoost classifier and
exploitation of Hidden-Markov-Models for trajectory
classification. The compound event approach is
demonstrated on detection of "dangerous occurrence of
the person on the platform edge" in underground
scenario. The methods process the video data from
standard low-resolution CCTV surveillance system.
The developed approaches are evaluated on real data
and applied in the real sites in underground scenarios.