12-11-2012, 06:22 PM
An abandoned object detection system based on dual background segmentation
An abandoned object detection system based on dual background segmentation IEEE 2009.pdf (Size: 491.3 KB / Downloads: 36)
INTRODUCTION
Recent years have seen a stark rise in terrorist attacks on
crowded public places such as airports, train stations and
subways, nightclubs, shopping malls, markets, etc. Many
surveillance tools have been employed in the fight against
terror. Although video surveillance systems have been in
operation for the past two decades, the analysis of the CCTV
footage has seldom ventured out of the hands of human
operators. Recent studies [1-3] have brought into fore the
limits to human effectiveness in analyzing and processing
crowded scenes, particularly in video surveillance systems
consisting of multiple cameras.
The advent of smart cameras with higher processing
capabilities has now made it possible to design systems
which can possibly detect suspicious behaviors (in general)
and abandoned objects (in particular). A number of
algorithms [5, 7, 8] have been suggested in the recent past to
deal with the problem of abandoned-object-detection. Due to
their dependence on complex probabilistic mathematics,
most of these algorithms have failed to perform satisfactorily
in real time scenarios. In addition, the other difficulty of
detecting an abandoned object under occlusion adds to the
overall complexity. Some proposed algorithms [4-5] have
dealt with partial occlusion (by moving people) but complete
or prolonged occlusion (by another object) has not yet been
tackled.
BACKGROUND SEGMENTATION
Numerous background subtraction methods are available
in the literature. The most popular being the ones based on
Gaussian mixture models, the first of which was proposed by
Friedman and Russell [12] and then modified by several
authors [13-14] to suit their specific needs. In this work, a
new background subtraction technique based on the
Approximate Median algorithm is developed. This method is
adaptive, dynamic, non-probabilistic and intuitive in nature.
Like the majority of other methods (for ex. [6]), we also use
pixel color/intensity information for background processing.
But instead of having one reference frame, we maintain two
different reference frames for self adaptability resulting in
less computation due to non-inclusion of any complex
mathematics. Moving crowd/objects, lighting changes and
unnecessary details like shadows, reflections on floors and
walls are filtered off efficiently with only stationary objects
remaining in the scene, thus leaving us with the prime
motive of ‘detecting abandoned objects’.
Algorithm
The proposed algorithm to separate background and
foreground in the incoming image is based on the
‘Approximate Median Model’ [6]. However, our technique
requires two reference background images, namely, ‘Current
Background’ and ‘Buffered Background’. This technique of
storing two backgrounds can be considered as a dual
background method. One of the interesting features of this
technique is that both the backgrounds are updated
dynamically. The first one is updated frequently while the
second one has a slower update rate.
The first frame of the incoming video is initialized
as ‘Current Background’. Subsequently, the intensity of each
pixel of this current background is compared with the
corresponding pixel of the next frame (after every 0.4
seconds). If it is less, then the intensity of that pixel of
current background is incremented by one unit, otherwise it
is decremented by one unit. In case of equality, the pixel
intensities remain unchanged. This way, even if the
foreground is changing at a fast pace, it will not affect the
background but if the foreground is stationary, it gradually
merges into the background
CONCLUSIONS
This paper presented an abandoned object detection
system based on a dual background segmentation scheme.
The background segmentation is adaptive in nature and
based on the Approximate Median Model. It consists of two
types of reference backgrounds, Current and Buffered
background, each with a different time interval. Blob
analysis is done on the segmented background and a
dynamic tracking algorithm is devised for tracking the blobs
even under occlusion. Detection results show that the
system is robust to variations in lighting conditions and the
number of people in the scene. In addition, the system is
simple and computationally less intensive as it avoids the
use of expensive filters while achieving better detection
results.