Abstract
The ability to quickly detect and locate stealthy radiological dispersal devices
(RDDs) allows authorities to disarm and remove the RDDs before they can be detonated.
Traditionally, the detection of RDDs was accomplished by using expensive and
cumbersome radiation sensors strategically located in the surveillance area. However,
with recent advancements in wireless technologies and sensing hardware, deploying
a large scale sensor network with small sensors is now becoming a reality. In this
dissertation, we study methods to detect and locate radiation sources quickly and
accurately using a network of sensors.
Localization of a single radiation source can be achieved by using three sensors
in a noise- and error-free environment. When both noise and errors are considered,
we present a closed-form solution that outperforms existing algorithms. When more
than three sensors are available, we present an efficient algorithm to exploit the additional
sensor data, in order to further improve the robustness and accuracy of the
localization.
To localize multiple sources in a sensor network, we propose a hybrid formulation
of a particle filter with a mean-shift technique, in order to achieve several important
features which address major challenges faced by existing multiple source localization
algorithms. First, our algorithm is able to maintain a constant number of estimation
(source) parameters even as the number of radiation sources K increases. Second,
our algorithm “learns” the number of sources from the estimated source parameters
instead of relying on expensive statistical estimations. Third, the presence of obstacles
may improve the localization accuracy of our algorithm. Unfortunately, the presence
of obstacles significantly degrades the accuracy of existing algorithms.
When no radiation source is present, the localization algorithms produce false
positives as the algorithms assume that a radiation source is present. We propose
the Localization Enhanced Detection (LED) method, that decides whether a source
with the estimated parameters is present or absent, using a close-to-minimal number
of measurements, while maintaining the false positive and false negative rates below
a specified level. We evaluate the LED method using simulation and testbed experiments,
and compare the effectiveness of the LED method with existing detection
methods.
We build a cross-platform, cross-language, and versatile software framework that
provides an abstraction for interfacing with sensors and supports building applications
on radiation source localization. The software framework implements various
localization algorithms that are ready to be deployed in an actual system. The components
in the software framework are loosely coupled and are general enough to
support application domains beyond radiation source localization. We demonstrate
the versatility of the software framework in building the Rapid Structural Assessment
Network.
