Cryptography Reference
In-Depth Information
9.1 Introduction
Intelligent video-based surveillance has become a very popular research topic
in recent years due to increasing crime. In the past, video-based surveillance
systems relied primarily on human operators to observe several, or dozens, of
monitors simultaneously. However, this kind of monitoring is not practical,
because a human operator cannot watch so many TV screens simultaneously
and observe everything that happens. Furthermore, unlike machines, human
operators become tired, especially after working long hours. Therefore, it is
essential to develop a smart real-time video-based surveillance system. Some
di
cult design issues must be taken into account. Firstly, the motion of a
human being is highly articulated, which makes the task of description very
di
cult. Secondly, in order to correctly identify an event, comparison of two
arbitrary motion sequences is indispensable.
In the past decade, extensive research have been conducted into surveillance-
related issues [1][16]. In [1], Wren et al. proposed a statistical background model
to locate people and utilized 2D image analysis to track a single person in complex
scenes with a fixed camera. Lipton et al. [2] proposed the use of a human operator
to monitor activities over a large area using multiple cameras. Their system can
detect and track multiple people and vehicles in crowded scenes for long periods.
Grimson et al. [3] established a multiple-camera environment to learn patterns
common to different object classes and detect unusual activities. Bobick et al. [4]
proposed a combination of a Hidden Markov Model and stochastic grammar
to recognize activities and identify different behavior based on contextual
information acquired by a static camcorder. Kanade et al. [5] proposed the
use of multiple sensors to detect and track moving objects. Haritaoglu et
al. [6] proposed a method for the detection and tracking of multiple people
and monitoring of their activities in an outdoor environment. Stauffer and
Grimson [7] focused on motion tracking in outdoor environment. They used
observed motions to learn patterns from different kinds of activities. In [8],
Zelnik and Irani proposed a non-parametric approach to characterize video
events by Spatio-Temporal Intensity Gradients calculated at multiple temporal
scales. Medioni et al. [9] used a set of features derived from multiple scales to
stabilize an image sequence. They proposed the extraction of the trajectories of
moving objects using an attribute graph representation. Davis and Bobick [10]
proposed an activity recognition method based on view-dependent template
matching. In [11], Davis et al. proposed the representation of simple periodic
events, such as walking, by constructing dynamic models based on computing
periodic patterns in peoples movements. Makris and Ellis [14] proposed to
automatically learn an activity-based semantic scene model based on motion
trajectories from video streams. In [15], Su et al. proposed the concept of motion
flow and used it to represent an event directly in a compressed domain. They
also proposed a scale and translation invariant matching algorithm to compare
an unknown event with database events.
