30-07-2012, 04:00 PM
Human Motion Detection and Tracking for Video Surveillance
HMD.pdf (Size: 491.91 KB / Downloads: 124)
Abstract
An Automated Video Surveillance system is presented
in this paper. The system aims at tracking an object in motion and
classifying it as a Human or Non-Human entity, which would help
in subsequent human activity analysis. The system employs a
novel combination of an Adaptive Background Modeling
Algorithm (based on the Gaussian Mixture Model) and a Human
Detection for Surveillance (HDS) System. The HDS system
incorporates a Histogram of Oriented Gradients based human
detector which is well known for its performance in detecting
humans in still images. Detailed analysis is carried out on the
performance of the system on various test videos.
INTRODUCTION
Automated Video Surveillance addresses real-time
observation of people and vehicles within a busy environment
leading to a description of their actions and interactions [1].
Technical issues include moving object detection and tracking,
object classification, human motion analysis, and activity
understanding.
Motivation and present state of research
A standard surveillance system attempts to recognize the
regions of interest in a video scene, i.e. the moving entities in
the scene. The classification of the moving entity forms a
critical part of the system, as the subsequent modules analyze
the moving entity based on whether it is a vehicle, a human or a
group of humans.
HUMAN DETECTION FOR SURVEILLANCE (HDS) SYSTEM
The algorithm description in the following section is based
on the Navneet Dallal and Bill Triggs’ Histogram of Oriented
Gradients (HOG) for Human Detection [4]. Their
implementation divides the image window into small spatial
regions called cells. Each cell accumulates a local 1-D
histogram of gradient directions or edge orientations of the
pixel values in the cell. The histogram entries combine to give a
unique representation for the image. Contrast normalization is
also carried out across the cells to give better invariance to
illumination changes.
Test Results
The INRIA-Person image database was used for training and
testing the Support Vector Machine. The database is available
for download at http://lear.inrialpes.fr/data. A total of 615
positive trainings samples were taken from the INRIA dataset
and 1200 randomly sampled regions (not consisting of humans)
were taken as the negative training set. The trained SVMLight
was tested on the test images provided in the INRIA-dataset and
achieved a recognition rate exceeding 90 percent. 288 positive
samples and 500 random negatives samples from the INRIA
dataset are used for testing. The results are given in Table 2.
The HDS system was also tested on a number of videos of
varying content. It was initially tested on a simple video
consisting of a human figure walking across the frame. The
system was able correctly detect the motion object and classify
it as human with an accuracy measure of 85.7% (Table 2). The
system was also tested on whether it could distinguish a human
person from an animal in a video sequence. An example of an
animal being tracked by the HDS system is shown in Fig. 6.
The gradient image clearly shows a marked difference between
the gradient arrangement of a dog and a human being (Fig. 4
and 5). This information is what enables the HDS system to
distinguish between the two classes.
CONCLUSION
A novel combination of Adaptive Background Modeling and
Histogram of Oriented Gradients was presented for tracking
and detecting human motion in a surveillance video sequence.
The Human Detection for Surveillance (HDS) system was
formulated, which could successfully classify a given image to
be as Human or Non-Human in nature. The system used a
Histogram of Oriented Gradients based approach for generation
of feature vectors. The feature vectors were classified into the
appropriate categories using a Support Vector Machine. The
integrated system was successfully tested on a number of
sample videos containing various moving entities. Analyses
were made on its performance and certain key issues were
identified to be kept in consideration while implementing the
system.