International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 4 (2018) pp. 1866-1873
© Research India Publications. http://www.ripublication.com
1866
A Novel Method to Improve Basic Background Subtraction Methods for
Object Detection in Video Surveillance System
Surender Singh
Research Scholar, School of Computer Science and Engineering,
University of Petroleum and Energy Studies, Dehradun, India.
Orcid Id: 0000-0002-7996-8588
Ajay Prasad
Associate Professor, School of Computer Science and Engineering,
University of Petroleum and Energy Studies, Dehradun, India.
Kingshuk Srivastava
Assistant Professor, School of Computer Science and Engineering,
University of Petroleum and Energy Studies, Dehradun, India.
Suman Bhattacharya
Project Head, IPR Management Services,
Tata Consultancy Services, Bhubneswar, India.
Abstract
This paper proposes a novel method for the improvement of
basic Background Subtraction (BGS) methods to detect moving
objects in video surveillance streams. The method is based on
Local Neighborhood Differencing (LND) in which instead of
finding a simple pixel to pixel difference between current frame
and background model, the average of the pixel neighborhoods
from the current frame and background model are subtracted to
entitle the pixel a background or foreground in the current frame
in order to find moving objects in video. The proposed method
has been tested on two basic methods; Adaptive Mean and
Adaptive Median methods of object detection using various
complex real time benchmarked scenarios. It is also compared
with classical statistical thresholding method. The results have
been measured in precision and recall metrics to register
improvement. The obtained results have confirmed the utility of
the method by increasing the robustness of the object detection
techniques in video surveillance for real time video analytic.
Keywords: Object Detection, Motion Detection, Background
Subtraction, Automated Video Surveillance, Adaptive Mean,
Adaptive Median, P-R Curves.
INTRODUCTION
Real-time video surveillance systems detect objects or
situations in video flow that represent a security threat and
trigger an alarm accordingly. These systems can be classified
into operator controlled, automated video surveillance and
intelligent video surveillance systems [19]. In operator
controlled surveillance system, the video stream is analyzed
manually; a person observes the video to determine if there is
any activity that requires an action. In the second approach, the
automated video surveillance system uses motion detection
techniques to determine response. An intelligent video
surveillance system is that which extract the relevant
information from generic motion accurately and issue actions.
Existing video surveillance systems take care about video
capture, store and transmission of video to remote places but
devoid of efficient object detection and analysis leaving these
functions exclusively to human operators for manual analysis
[6]. Therefore, there is an urgent need of a surveillance system
which is fast, efficient and accurate.
There are several categories of object detection methods out of
which the background subtraction is most popular and
traditionally used category [18]. In this category, there are
robust methods such as Kernel Density Estimation and
Histogram Detection which provide reliable detection but these
are also slow and less useful for real time analytics. This
category also includes some basic methods like Frame
Differencing, Adaptive Mean, Adaptive Median methods which
are fast but do not provide good object detection results. Our
research is targeted to improve these basic BGS methods for
object detection.
This paper is organized as follows: Section 1 identifies the
background on the need of modification in object detection
paradigm to make it more robust and useful for real time
scenarios. Section 2 presents the related works which made
efforts to improve the basic methods. Section 3 describes the
proposed methodology. Section 4 tabulates and compares the
results obtained with the proposed methods and other past
improvements in basic methods. Section 5 discusses the results
which is followed by conclusion and future scope in section 6.
RELATED WORKS
Most of the improvements made in the past in basic background
subtraction methods revolves around proposing a threshold
which is effective and adaptive in different situations or
scenarios [1]. Many statistical measures such as mean, median,
deviation, outliers and variance from mean are used to define a