22-06-2012, 02:01 PM
Probabilistic Framework for Person Tracking on Embedded Distributed Smart Cameras
[attachment=25074]
INTRODUCTION
Networks of multiple smart cameras can be used to efficiently
implement features like detection accuracy, faulttolerance
and robustness that wouldn’t have been possible with
a single camera in such applications like surveillance or smart
environments. This paper focuses on a distributed multi camera
visual tracker system where each camera’s field of views (FoV)
does not overlap with others.
The effectiveness of Bayesian algorithm for tracker system
has been successfully represented in several systems.
A multi-person distributed tracking system was proposed by
Mensink[1]. The appearance and spatial-temporal features
were applied to a probabilistic model to find the correspondence
between a person’s identity and an observation.
Javed et al. [2] proposed an approach for establishing object
correspondence across non-overlapping cameras.
PROBABILISTIC GENERATIVE TRACKER MODEL
This section presents our probabilistic visual tracking model
which uses different features of an individual in a typically
disjointed camera setup. The probabilistic framework is responsible
for fusion of different extracted features for the
final inference. The features are categorized to appearance,
environmental, and motion models.
The appearance feature expresses the color histogram of
cloths of a person. Different cameras with various inherent
sensor characteristics and also under different light conditions
acquire slightly different color expressions. Lets suppose each
sample of this feature comes from a Gaussian distribution.
SYSTEM ARCHITECTURE
A. Data exchange among multiple cameras
In order to preform the data exchange among cameras, we
use the middleware technology. The middleware dwells between
the vision application layer and the underlying operating
system, network protocol stack, and hardware. In order to meet
the required communication constraints dictated by the upper
application layer (i.e, distributed person tracker), we proposed
a novel asynchronous event exchange mechanism in [12].
EXPERIMENTS
We mounted three video cameras apart in our lab, and
corridor of the building, which were attached to the tripods and
walls. We used two interconnected suite of PCs with Logitech
Webcam C500. OpenCV was used for implementation of
computer vision and machine learning (ML) algorithms such
as EM and SVM. TAO[15] was applied as the middleware
for exchanging the abstract information between cameras. Our
developed embedded FPGA-based smart camera system[16]
was exploited as the third node. It includes the Virtex4 (Xilinx)
FPGA device.
CONCLUSION AND FUTURE RESEARCH
Through this paper, we demonstrated the concept of a
distributed person tracker system. The tracker leverages a
probabilistic modeling based on the linear interpolation to
combine the person’s observed features in a non-overlapped
installed camera setup.
[attachment=25074]
INTRODUCTION
Networks of multiple smart cameras can be used to efficiently
implement features like detection accuracy, faulttolerance
and robustness that wouldn’t have been possible with
a single camera in such applications like surveillance or smart
environments. This paper focuses on a distributed multi camera
visual tracker system where each camera’s field of views (FoV)
does not overlap with others.
The effectiveness of Bayesian algorithm for tracker system
has been successfully represented in several systems.
A multi-person distributed tracking system was proposed by
Mensink[1]. The appearance and spatial-temporal features
were applied to a probabilistic model to find the correspondence
between a person’s identity and an observation.
Javed et al. [2] proposed an approach for establishing object
correspondence across non-overlapping cameras.
PROBABILISTIC GENERATIVE TRACKER MODEL
This section presents our probabilistic visual tracking model
which uses different features of an individual in a typically
disjointed camera setup. The probabilistic framework is responsible
for fusion of different extracted features for the
final inference. The features are categorized to appearance,
environmental, and motion models.
The appearance feature expresses the color histogram of
cloths of a person. Different cameras with various inherent
sensor characteristics and also under different light conditions
acquire slightly different color expressions. Lets suppose each
sample of this feature comes from a Gaussian distribution.
SYSTEM ARCHITECTURE
A. Data exchange among multiple cameras
In order to preform the data exchange among cameras, we
use the middleware technology. The middleware dwells between
the vision application layer and the underlying operating
system, network protocol stack, and hardware. In order to meet
the required communication constraints dictated by the upper
application layer (i.e, distributed person tracker), we proposed
a novel asynchronous event exchange mechanism in [12].
EXPERIMENTS
We mounted three video cameras apart in our lab, and
corridor of the building, which were attached to the tripods and
walls. We used two interconnected suite of PCs with Logitech
Webcam C500. OpenCV was used for implementation of
computer vision and machine learning (ML) algorithms such
as EM and SVM. TAO[15] was applied as the middleware
for exchanging the abstract information between cameras. Our
developed embedded FPGA-based smart camera system[16]
was exploited as the third node. It includes the Virtex4 (Xilinx)
FPGA device.
CONCLUSION AND FUTURE RESEARCH
Through this paper, we demonstrated the concept of a
distributed person tracker system. The tracker leverages a
probabilistic modeling based on the linear interpolation to
combine the person’s observed features in a non-overlapped
installed camera setup.