42

3. Orthogonal Projection and Kalman Filter

we have
....

Pk,k ==(1 - GkCk)Pk,k-l(I - GkCk)

T....

+ (I -

..

GkCk)Pk,k-l(GkCk)

==(1 - OkCk)Pk,k-l .

(3.23)

Therefore, combining (3.13), (3.16), (3.18), (3.21), (3.22) and

(3.23), together with
Po,o == II x o - xOloll~ == Var(xo) ,

(3.24)

we obtain the Kalman filtering equations which agree with the

ones we derived in Chapter 2. That is, we have xklk == xklk' xklk-l ==
xklk-l and Ok == Gk as follows:
Po,o == Var(xo)
Pk,k-l == Ak-lPk-l,k-lAl-l + fk-lQk-lfr-l
Gk == Pk,k-l C -:' (CkPk,k-1C-:' + Rk)-l
Pk,k == (I - GkCk)Pk,k-l

== E(xo)
xklk-l == Ak-1Xk-llk-l
xklk == xklk-l + Gk(Vk - CkXk1k-l)
k == 1,2, ....
xOlo

(3.25)

Of course, the Kalman filtering equations (2.18) derived in

Section 2.4 for the general linear deterministic/stochastic system
Xk+l
{

Vk

== AkXk + BkUk + rk~k

== CkXk + DkUk +!1k

can also be obtained without the assumption on the invertibility

of the matrices Ak, VarC~k,j)' etc. (cf. Exercise 3.6).

3.5 Real-Time Tracking

To illustrate the application of the Kalman filtering algorithm described by (3.25), let us consider an example of real-time tracking.
Let x(t), 0 :::; t < 00, denote the trajectory in three-dimensional
space of a flying object, where t denotes the time variable (cf.
Fig.3.1). This vector-valued function is discretized by sampling
and quantizing with sampling time h > 0 to yield
Xk

x(kh),

== 0,1,.

3.5 Real-Time 'fracking

43

Fig. 3.1.

, ", - -

-~
x(t)

"'-,
I
I

x(O)

For practical purposes, x(t) can be assumed to have continuous

first and second order derivatives, denoted by x(t) and x(t), respectively, so that for small values of h, the position and velocity
vectors Xk and Xk ~ x(kh) are governed by the equations

1h2 ..

~k+l

= ~k + h~k +"2

Xk+l

= Xk + hXk ,

Xk

where Xk ~ x(kh) and k = 0,1,. In addition, in many applications only the position (vector) of the flying object is observed
at each time instant, so that Vk = CXk with C = [I 0 0] is
measured. In view of Exercise 3.8, to facilitate our discussion we
only consider the tracking model

(3.26)

44

3. Orthogonal Projection and Kalman Filter

to be zero-mean Gaussian white noise sequences satisfying:

E(~k)

= 0,

E("1k) = 0,

E(~k~;) = Qk 6kl,
E(xo~;) = 0,

E("1k"1l) = rk 6kl,
E(Xo"1k) = 0,

where Qk is a non-negative definite symmetric matrix and rk > 0

for all k. It is further assumed that initial conditions E(xo) and
Var(xo) are given. For this tracking model, the Kalman filtering
algorithm can be specified as follows: Let Pk := Pk,k and let P[i, j]
denote the (i, j)th entry of P. Then we have
Pk,k-l[l,l] = Pk-l[l, 1]

+ 2hPk-l[1, 2] + h2Pk_l[1, 3] + h 2Pk-l[2, 2]

h4

+ h 3 Pk-l[2, 3] + 4Pk-1[3, 3] + Qk-l[l, 1],

P k,k-l[1,2] = Pk,k-l[2, 1]
3h 2

= Pk-l[l, 2] + hPk-l[l, 3] + hPk-l[2, 2] + TPk-1[2, 3]

h3

Pk,k-l[2,2] =
Pk,k-l[1,3]

+ 2Pk-1[3, 3] + Qk-l[l, 2],

2
Pk-l[2, 2] + 2hPk-l[2, 3] + h Pk-l[3, 3] + Qk-l[2, 2],
Pk,k-l[3, 1]

h2
= P k- 1[1, 3] + hPk- 1[2, 3] + 2Pk-1[3, 3] + Qk-l[l, 3],

= Pk,k-l [3,2]
= Pk-l[2, 3] + hPk-l[3, 3] + Qk-l[2, 3],
Pk,k-l [3,3] = Pk-l [3,3] + Qk-l [3,3] ,
Pk,k-l [2,3]

with Po,o = Var(xo) ,

Exercises

45

(3.27)

with :Kala = E(xo).

Exercises
3.1. Let A =f. 0 be a non-negative definite and symmetric constant
matrix. Show that trA > o. (Hint: Decompose A as A = BB T
with B =f. 0.)
3.2. Let
j-I

ej

Cj(Xj -

Yj-I) = C j

(Xi -

Pi-l,iVi) ,

'1.=0

where Pj-I,i are some constant matrices. Use Assumption 2.1

to show that
for all /!, ? j.
3.3. For random vectors

define

WO,"', W r ,

Y(Wo,"', w r )
r

Po, ... 'Pr' constant matrices}.

y= LPiWi,
i=O

Let

j-I

Zj

= Vj

Cj

Pj-l,iVi

i=O

be defined as in (3.4) and

ej

3.4. Let

Ilzjll-lzj'

Show that

j-I

Yj-I

=L

Pj-l,iVi

i=O

and

j-I
Zj

Vj -

Cj

Pj-l,iVi .

i=O

Show that
j

= 0,1," . ,k - 1.