15-09-2016, 12:37 PM
1454603648-downloadfile1.doc (Size: 222.76 KB / Downloads: 5)
INTRODUCTION
The task of automatically detecting frames with
anomalous or interesting events from long duration video
sequences has concerned the research com-munity in the last
decade. Event, and especially anomaly detection in crowded
scenes is very important, e.g. for security applications, where
it is difficult even for trained personnel to reliably monitor
scenes with dense crowds or videos of long duration.
Numerous methods have been proposed to assist in this
direction.
The analysis of motions and behaviors in crowded scenes
constitutes a challenging task for traditional computer vision
methods, as barriers like occlusions, varying crowd densities and
the complex stochastic nature of their motions are difficult to
overcome. Computational cost is one more complicating factor,
as it has to be kept within reasonable limits. In many practical
situations, it is crucial to analyze crowded scenes in real time, or
at least as fast as possible, considering the fact that security
personnel should act quickly if something seems to be “not as
usual.” Furthermore, the ambiguity of the term “anomaly” sets its
own limitations in our effort to identify it, as there is no
commonly accepted definition, and it varies significantly
depending on the given scenario. This means that an “anomaly”
pattern in one video sequence may often be part of the “normal”
pattern of another. In order to address these issues, we define as
“anomalies” the events that display a low probability of occurring
based on earlier observations.
Related Work:
2.1 STATE OF THE ART
Even though significant research has taken place on event
and anomaly detection from static cameras [2], [3] the
majority of these works address non-crowded scenes, where
detailed visual information can be exploited for each
individual. However, real-world surveillance scenarios often
involve crowds of people or dense traffic, where such
information cannot be easily extracted with traditionally used
methods. Therefore, a number of different approaches have
been proposed to handle these situations.. Existing methods
can be classified in two main categories: those that use only
motion information to detect an abnormality in the scene, and
those that use both appearance and motion information to
describe the scene dynamics.
In the first category, Wu et al. [8] use chaotic dynamics in
particles’ representative trajectories as a means to build a
model capable of locating an outlier that moves with a
different pattern. Even though this method works for very
dense videos where a global motion pattern exists