Ian Reid

Hilary Term 2003

Copies from http://www.robots.ox.ac.uk/ ian/Teaching/Estimation
Bugs to ian.reid@eng.ox.ac.uk

APPLIED ESTIMATION I

1. Joint probabilities for the discrete random variables

 are given by


 





0.2 0.07 0.06
 0.09 0.1 0.01
 0.02
0.06
0.12
0.08
0.09
0.1

(a) Check the normalisation of the joint distribution


   



(b) Compute the marginal probability distributions for
(c) Compute the conditional distribution 

 .

(d) Use Bayes’ rule to compute 

   using only information in the previous 2
parts; then check the answer by direct calculation from the table above.

   " !
#%$ '& $! )(*+,

2. (a) If the independent continuous variables are normally distributed with means and
variances , derive the distribution of their sum (Hint: Consider -/.0!
the moment generating function (Characteristic function) of the random variables)

2 
(b) Now suppose that those variables are no longer independent but have correlation co-
1
3     ! 54
efficient . Find the covariance and information matrices of the vector variable
. Write down the joint density function including the normalisation constant.
3. (a) Find the conditional density function 678 !     for the problem above, including the

9*:  !   ; and the conditional variance <=:  !    ; ?

normalisation constant.
(b) What are the conditional mean



4. The data of question
to measure a state .
1 could be taken as characterising a noisy sensor with output , intended

> has been taken, what is the MLE for
?

(a) Given that a measurement
(b) Now, if conditional probabilities 
(
@?A are as in part a) but prior probabilities
are as follows:


BCDFEG
H 

CDIE JF 
K> CDIE
find the MAP estimate for .


5. Independent sensors have outputs LM LN!
 %() ,
measuring a state  which are unbiased and
normally distributed with standard deviation $
& O(PQ
.
(a) Derive the maximum likelihood estimate of  given single measurements LM LN! from

each sensor.

1
 (b) If also the state itself is known to have a prior distribution that is normal R   ' & S!  what
is the posterior density for ? 
(c) If, instead, the prior density is given by the exponential distribution
^]`_ D
6O3 *UTWV*D XZYA[ 5\ V  cQd'ebXga f'h ]ji X
and only a single sensor reading LM is made, what is the MAP estimate for  ?
(d) Suppose in part b) that
 kgDIO& S >DIE Il&  >
 & ! DIE 
and that the sensor readings are
L  gDIE J L !  N
HE DFE

What is the MAP estimate for , based on these data?
LM LN!
(e) Apply validation tests to and and show that one sensor fails and the other passes.

(f) What now is the MAP estimate for in b), given only the validated sensor reading?

m\ 
6. Consider a surveillance system that uses a video camera to track vehicle travelling down a
steep ramp into a car park. Vehicle position at timestep is cm and from the m 2%n
th to the
m
 DI'DIE` m 54
th time step the position of a vehicle is assumed to change by a random displacement whose
mean is and whose variance is
o DIE`pDFEq
DIE`pDFE sr
(a) If the initial mean and covariance of the vehicle’s position (at time m D ) are
o gDD o BD gD
gDD@rutQvIw BDx
MD@r
what is the probability distribution after m
kBDQD time-steps?
(b) Now a two-dimensional measurement y of vehicle position is made that is unbiased and
 3z DF'{DQD 4 what
with a random error that is normal and isotropic — ie such that its standard deviation in
any direction is constant — and the value of that s.d. is 10 cm. If y

gD timesteps, a second measurement is to be made. What is the effective

now is the posterior distribution for the vehicle’s position?
(c) After a further
prior density for 2 , just before making the new measurement?
(d) The second measurement LB!
|}DD is made by a different sensor that is one-dimensional,
measuring distance in the ~ direction only. It is a biased sensor that under-reads by cm

and has random error with s.d. 5cm. What now is the posterior density of the position?
(e) Write a matlab function that would plot each of these distributions as covariance ellipses
(and plot them, if you like).


n '  n *   DF,D€  ZN 

HgQE €  I,
HEGJQ
7. Data

is given in which the variance associated with the n is


& n! & ! 
constant for all m .

2
 (a) Use the recursive least-squares algorithm to fit a linear function to the data.
(b) Compare your result with the “batch” algorithm that uses the pseudo-inverse of the “Van-
dermonde” matrix (see Engineering Computation 2nd year notes).

Answers
2. b) 673‚ Cƒ…„  X€YH[ \ ! 8 2†\2 ‡  47ˆ 8 2l\2 ‡  with ƒK
}‰%&  & !Š  \‹1 ! and

ˆ    o NŒM& Œ}! & & \}Œ}1 & Œ}! &  & ! r

\‹1 ! \1  ! !
 R  "‡ ! ,&-!Ž  with
3. a) "!  
&
"‡ !  #%!/. & ! 1"3 \#O  tQvIw & !Ž >& ! Š  \‹1 ! E

4. a) 

 ; b)
K 
.
& ! ]‘_ &!
  +T MLD /\ V  Mc d'eL}=ZX f,a h q] V i X 
5. c)

6.
a) Normal R 2 ‡   ‡  with

2 ‡ o {BDQDDJ@r   ‡  o
MM
Dx
QD
DxJQD0r
b) Normal R  2     with
o |HE J o ZIE  QE 
2   zQ{ D
Hz EG
|‹r    Q QE x FE 0r
c) Normal R  2 ‡ Ž   ‡ Ž  with

Ž2 ‡  o z|MDA| |AE z ’E J|Q| r   ‡  o BB JF

FEE x
 N
HE 
QIE  r
d) Normal R  2  Ž“  Žq with
o |AEqN| o g IE
F”JFE`}|
2  Ž  z}| DJFE J|‹r   Ž  HJ EqN| gIEGJQJ@r
7. a)
•DFE DQ . ID E  M
.