22.01.2015 Views

Military Communications and Information Technology: A Trusted ...

Military Communications and Information Technology: A Trusted ...

Military Communications and Information Technology: A Trusted ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Chapter 3: <strong>Information</strong> <strong>Technology</strong> for Interoperability <strong>and</strong> Decision...<br />

287<br />

in the fusion center. In particular, it is possible to modify all local tracks such that<br />

they become decorrelated. The cross-correlations occur because of a common evolution<br />

covariance in every prediction step [5]. This is due to the fact that all sensors<br />

sites observe the same target. In [6] it is shown that a central Kalman filter can be<br />

calculated in a distributed manner by decorrelating the local tracks. This is achieved<br />

by calculating the global estimation error covariance of the system. The proposed<br />

distributed Kalman-type processing scheme makes essentially use of the fact that<br />

the sensor measurements do not enter into the update equation for the estimation<br />

error covariance matrices. In particular, the covariance matrices of all sensors<br />

can be calculated at each individual sensor site without any further communication<br />

(provided the relevant parameters of all sensors are known at each sensor site).<br />

For the sake of notational simplicity, let all S sensors be equally aligned <strong>and</strong><br />

synchronized with the same data update rate. However, these assumptions are not<br />

essential <strong>and</strong> can well be relaxed. Furthermore, we assume the measurement error<br />

covariance matrices { R<br />

s }<br />

S<br />

k s <br />

of all individual sensors to be known to each local sensor<br />

processor. As mentioned above, the proposed distributed processing scheme aims at<br />

establishing decorrelated local tracks, such that fusing them yields exactly the result<br />

of central Kalman filter processing. Let us assume, a set of decorrelated local tracks<br />

at time are given which have processed all sensor data up to time Then, as all<br />

densities involved are Gaussians, we can write the fused track as the following product:<br />

S<br />

k s s<br />

l lk |<br />

N (<br />

l; xlk |<br />

,<br />

lk |<br />

s1<br />

px ( | Z) c x P)<br />

,.<br />

(21)<br />

In the sequel, as well as in most applications, it is unnecessary to calculate<br />

the normalization constant c<br />

lk<br />

explicitly. By virtue of the factorization lemma for<br />

Gaussians, this product representation can be transformed into a single Gaussian:<br />

S<br />

k s s<br />

l lk |<br />

N<br />

l; lk |<br />

,<br />

lk |<br />

s1<br />

px ( | Z) c ( x x P)<br />

(22)<br />

N ( x ; x , P ),<br />

(23)<br />

l lk | lk |<br />

with an expectation vector x<br />

lk<br />

<strong>and</strong> a covariance matrix P<br />

lk<br />

obtained by fusing x<br />

lk<br />

<strong>and</strong> P s , s1, , S,<br />

according to following equations:<br />

lk<br />

S<br />

1 s 1<br />

lk |<br />

Plk<br />

|<br />

s1<br />

P<br />

(24)<br />

S<br />

s<br />

s<br />

lk | lk | lk | lk |<br />

s1<br />

x P ( P x ).<br />

(25)<br />

Convex combinations of this type are fundamental in almost all data fusion<br />

applications (see e.g. [7]). As previously stated, under conditions where Kalman

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!