Evidence Filtering In Sensor Networks

This work provides an ultra-efficient way of selectively fusing multiple data streams generated from many heterogeneous sources. It has the capability of solving the "needle in a hay stack" problem.

Many factors contribute to data imperfections in distributed decision making environments. When large networks of inexpensive, multiple sensor modalities are employed for detection purposes, their reliability, resolution, sensitivity, sampling frequency, sensor proximity to the detected events, and background noise can cause significant inaccuracies in gathered information. Indicators derived from databases or subjective expert opinions used to complement the sensor data may also contribute to such imperfections. Moreover, often the situation under observation is inherently uncertain. Prior information or conditional probability distributions are not available and improper initial assumptions or interpolations can also weaken the integrity of the decision making process.

Surveillance applications often call for observing sensor data and various other indicators spanning over multiple modalities. Here the term multiple modalities refer to different types of measurements, i.e., metal from a vehicle detected using a magnetometer, an indication of suspicious activity using a database etc. Gathering such data over time and making inferences based on the frequency characteristics of certain events can sometimes uncover a key piece of information. For example, a periodically occurring pattern of an event characterized by a particular set of sensor modalities may indicate an imminent security threat in a homeland security application. Two main issues need to be addressed in this context:

(a) How can we model imperfect data from multiple sensor modalities during information processing?
(b) How can we make direct inferences on the frequency characteristics of events of interest?

In this research, we integrate Dempster-Shafer (DS) belief theory with discrete time filtering techniques to address these two issues. The novel Evidence Filtering method presented here is capable of fusing temporally ordered information from multiple sensor modalities to directly infer on the frequency characteristics of events. To our knowledge, no single strategy capable of providing inferences in the frequency domain of events based on data from multiple sensor modalities is yet available.

The advantage of using DS theory to model evidence lies in its ability to conveniently represent a wide variety of data imperfections. It has been extensively used in surveillance and security applications in the past, and provides an excellent framework to model imperfect data derived from multiple sensor modalities. Moreover, this approach is ideal for the present context since it is directly extendable to accommodate heterogeneous sources. The advantages in DS theoretic methods become evident when the assumptions typical of a Bayesian approach (e.g., conditional independence, availability of priors, etc.) are difficult to justify.

To see the resume of the expert associated with this case study, see the link below.