0% found this document useful (0 votes)
67 views9 pages

Baye Sian S Tate E Stim Atio N

This document discusses Bayesian state estimation and filtering techniques such as the Kalman filter and particle filters. It provides background on the mathematical tools and notation used for probabilistic state estimation of dynamic systems given sensor measurements over time.

Uploaded by

hemu07121990
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
67 views9 pages

Baye Sian S Tate E Stim Atio N

This document discusses Bayesian state estimation and filtering techniques such as the Kalman filter and particle filters. It provides background on the mathematical tools and notation used for probabilistic state estimation of dynamic systems given sensor measurements over time.

Uploaded by

hemu07121990
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Chapter 2

Baye sian S tate E stim atio n


Most of th e localiz ation , m appin g an d S L AM approach es h ave a probabilistic form u lation . In th is ch apter, we revise th e B ayesian state estim ation fram ework, an d in particu lar th e ltering p ro b lem , in ord er to provid e th e read er with backg rou n d on th e u sed m ath em atical tools, an d for in trod u cin g th e n otation . T h e lterin g problem can be ex pressed as estim atin g th e state x of a d yn am ic d iscrete system (see F ig u re 2 .1 ), g iven : th e an alytical kn owled g e of th e state tran sitio n fu n c tio n f t an d th e statistical kn owled g e of th e state n o ise w t , th e an alytical kn owled g e of th e o u tp u t fu n c tio n h t an d th e statistical kn owled g e of th e o b serv atio n n o ise vt , a realiz ation of th e system ou tpu t z1:t u p to tim e t. A probabilistic lter for a d yn am ic system is a m ath em atical tool, wh ose g oal is to estim ate a d istribu tion of th e possible system state h istories g iven th e m easu rem en ts: p(x0 :t | z1:t ) (2 .1 )
vt ht wt ft xt1 xt zt

F ig u re 2 .1 : G en eric S ystem Mod el. T h e block is a on e step d elay.

Bayesian S tate E stim atio n

15

Here x0:t are the system states from the in stan t 0 to the in stan t t, while z 1:t is the history of m easu rem en ts z from 1 to t. In m ost of the cases, on e is in terested in evalu atin g on ly the m arg in al d istribu tion of the cu rren t state g iven the observation s: p(xt |z1:t ) (2 .2 )

in stead of the fu ll state history. S everal on lin e an d o lin e techn iq u es for solvin g the lterin g problem have been proposed [3 5 ; 2 1; 1; 5 4 ]. M ost of them rely on the assu m ption that the process bein g observed is M arkovian . A process is M arkovian if the cu rren t m easu rem en t is in d epen d en t from the past on es, g iven the cu rren t state: p(zt |z1:t1 , xt ) = p(zt |xt ). In the con tex t of the S L AM problem , for the M arkov assu m ption to hold , n o m ovin g objects u n kn own to the robot can popu late the en viron m en t. T his im poses obviou s restriction s to the application d om ain s. However, in m od erately d yn am ic en viron m en ts m ost of the techn iq u es proposed in this section have shown to work. In case of big violation s of the M arkov assu m ption , a typical approach con sists in pre-processin g the lter in pu t in ord er to skip the sen sor read in g s g en erated by d yn am ic objects. In the rest of this chapter we d escribe a wid e ran g e of u sefu l tools for B ayes lterin g , which is a fram ework that can be u sed for estim atin g the state of M arkovian system s. In particu lar we will d escribe: T he K alm an F ilter (K F ), which is an ex act, closed form lter workin g with lin ear system s, a ected by z ero m ean G au ssian n oise. T he Particle F ilters (PF ), which are M on te C arlo m ethod s su itable for the state estim ation of n on lin ear n on G au ssian d yn am ic system s.

2.1

B a ye s Fra m e w o rk

N ow we shortly presen t the key lterin g eq u ation s, an d form aliz e the problem . L et p(z |x) be the observation m od el , that is the d en sity of the m easu rem en t z , g iven that the system state is x, an d let p(x t |xt1 ) be the evolu tion m od el1 . If the M arkov assu m ption hold s, the posterior of the state chain u p to tim e t
1 In th e fo llo w ing se ctio ns, th e e v o lu tio n m o d e l is a lso re fe rre d to a s motion mod e l, since it is u se d fo r d e scrib ing th e ch a ng e o f th e ro b o t sta te a fte r m o tio n.

Bayesian S tate E stim atio n is p(x0:t |z1:t ) = = =

16

p(zt |x0:t , z1:t )p(x0:t |z1:t1 ) [by M ark ov assu m ption ] p(zt |z1:t1 ) p(zt |xt )p(x0:t |z1:t1 ) p(zt |z1:t1 ) p(zt |xt )p(xt |xt1 ) p(x0:t1 |z1:t1 ) (2 .3 ) p(zt |z1:t1 )

If on e is in terested in estim atin g the cu rren t state distribu tion , the lterin g eq u ation becom es the followin g : p(xt |z1:t ) = = = p(zt |xt )p(xt |z1:t1 ) p(zt |z1:t1 ) p(zt |xt ) p(xt |xt1 )p(xt1 |z1:t1 )d xt1 p(zt |z1:t1 ) p(zt |xt ) p(xt |xt1 )p(xt1 |z1:t1 )d xt p(zt |z1:t1 , xt )p(xt |z1:t1 )d xt p(xt |xt1 )p(xt1 |z1:t1 )d xt1 (2 .4 )

= p(zt |xt )

here is a n orm aliz ation factor en su rin g that E q . 2 .4 correctly represen ts a probability distribu tion . U su ally, the evalu ation of E q . 2 .4 is don e in two steps: p red ic tio n an d u p d ate. In the prediction step, the resu lt of the state tran sition from x t1 to xt is com pu ted. In the u pdate step, the last observation z t is in corporated in the previou sly com pu ted probability den sity. Referrin g to E q . 2 .4 , on e can arg u e that the predict step con sists in com pu tin g the in teg ral term . T he u pdate step is perform ed by weig htin g the predicted belief p(xt |xt1 )p(xt1 |z1:t1 )d xt1 with the last observation lik elihood p(z t |xt ). T hese two steps can be fou n d in all of the lters described in the rem in der of this chapter. B ayes lterin g in this form is ex act an d can be u sed on an y k in d of system for which the M ark ov assu m ption holds. U n fortu n ately, in the above eq u ation s there are som e in teg ration s over the state space. In m an y cases the state space is hig h dim en sion al, an d the B ayes lterin g can n ot be directly im plem en ted. For in stan ce, in the S L AM problem the dim en sion of the system state is the su m of the robot location dim en sion an d the m ap space dim en sion , which can easily be in the order of the hu n dreds or thou san ds. A straig htforward evalu ation of E q . 2 .4 wou ld req u ire an in teg ration over the en tire state space. For this reason approx im ated techn iq u es are n eeded.

Bayesian S tate E stim atio n

17

2.2

K a lm a n F ilte r

The K alman F ilter (K F ) [3 5 ; 7 2 ] is an ex ac t lter that c an be d erived d irec tly by E q u atio n 2 .4 u n d er the assu mptio n s that the system is lin ear an d the n o ise is G au ssian . U n d er these lin earity hypo theses the system c an be d esc ribed by xt = Ft xt1 + wt zt = Ht xt + vt The system n o ise wt N (0 , wt ) an d the o bservatio n n o ise vt N (0 , vt ) are z ero mean n o rmally d istribu ted . The key ad van tag e o f the K alman F ilter is that it represen ts the d istribu tio n s in c lo sed fo rm, in terms o f mean s an d c o varian c e matrix . The u pd ate o f the K alman lter c an be c arried o u t in the time o f a matrix mu ltiplic atio n (O(n 3 ), where n is the state d imen sio n ). The iterative alg o rithm o f the lter is the fo llo win g : pred ic t: xt = Ft xt1 u pd ate:
T + vt Kt = t Ht Ht t Ht 1

t = Ft t1 FtT + wt

xt = xt + Kt zt Ht xt

t = ( I Kt H t ) t

U n fo rtu n ately, in the mo bile ro bo t d o main , the evo lu tio n mo d el, as well as the o bservatio n mo d el are n o n lin ear, thu s the n o ise c an n o t be c o n sid ered G au ssian . H o wever, fo r mild evo lu tio n laws, a n o n lin ear ex ten sio n c an be u sed : the E x ten d ed K alman F ilter (E K F ) [7 2 ], in which lo c al lin eariz atio n s o f the state tran sitio n fu n c tio n f an d the o bservatio n mo d el h are perfo rmed . The ex ten d ed K alman lter alg o rithm c an be ex pressed as pred ic t: xt = ft (xt ) u pd ate:
T + vt Kt = t Ht Ht t Ht 1

t = Ft t1 FtT + wt

xt = xt + Kt zt ht (xt )

t = (I Kt Ht ) t

here Ft = x ft |xt an d Ht = x ht |xt . The key limitatio n s in the u se o f ex ten d ed K alman lter lies in the stro n g assu mptio n s that have to be d o n e o n the estimated system, n amely: G au ssian n o ise, an d lin eariz ability. In mo st o f the ro bo tic systems u sed fo r lo c aliz atio n an d S L A M , the u n c ertain ty is n o t ex pressible, n o r appro x imable as a G au ssian d istribu tio n , bein g mu lti mo d al an d irreg u larly shaped . W hen mo re mo d es are

Bayesian S tate E stim atio n

18

present in a d istribu tion, d ealing with mu ltiple hypotheses is need ed , while the K alman F ilter work s on their mean. In su ch a situ ations, its u se is prone to failu re. M oreover, the lineariz ation of the system can introd u ce some systematic error in the estimate. F inally, some systems cannot be lineariz ed (being their 1st ord er d erivatives nu ll), thu s the ex tend ed K alman F ilter cannot be applied . In these contex ts, a second ord er ex tension to the K alman lter: the U nscented K alman F ilter (U K F ) has been proposed in [7 0]. W hile the U K F in g eneral behaves better than the K alman lter the hypotheses of G au ssian noise is still req u ired to hold . D espite the above ou tlined limitations, the K alman F ilter is one of the most u sed tools in localiz ation and S L AM , d u e to its simplicity. M oreover, when the u nd erlying hypotheses hold , it ex hibits a strong converg ence rate if compared with other ltering techniq u es.

2.3

Pa rtic le F ilte rs

A particle lter is a B ayes lter that work s by representing a probability d istribu tion as a set of samples (particles): 1 x(i) (x). (2 .5 ) p(x) N
i

where x(i) (x) is the impu lse fu nction centered in x (i) . T he d enser are the samples x(i) in a reg ion, the hig her is the probability that the cu rrent state falls within that reg ion. In principle, in ord er to maintain a sampled representation of the feasible (i) system state histories, one shou ld d raw the samples {x 0 :t } form the probability d istribu tion p(x0 :t | z1 :t ) of the cu rrent state g iven the observation history. S u ch a d istribu tion is in g eneral not available in a form su itable for sampling . H owever, the Im p o rtan c e S am p lin g (IS ) principle ensu res that if one can: evalu ate point wise and d raw samples from an arbitrarily chosen importance fu nction (x0 :t | z1 :t ), su ch that p(x0 :t | z1 :t ) > 0 (x0 :t | z1 :t ) > 0, and evalu ate point wise p(x0 :t | z1 :t ), then it is possible to recover a sampled approx imation of p(x 0 :t | z1 :t ) as p (x0 :t | z1 :t )
i (i)

w(i) x(i) (x0 :t ).


0 :t

(2 .6 )
(i)

H ere {x0 :t } are samples d rawn from (x0 :t | z1 :t ) and wt

p(x0 :t |z1 :t ) (x0 :t |z1 :t )


(i)

(i)

is the

importance weig ht related to the ith sample that tak es into accou nt the mismatch among the targ et d istribu tion p(x t | z1 :t ) and the importance fu nction.

Bayesian S tate E stim atio n

19

a)

b)

c)

Figure 2 .2 : T he Im portance S am pling principle. In picture a) the goal d istribution and som e sam ples d rawn from it are d epicted . T he m ore d ense are the sam ples in a region, the higher is the probability d ensity in that region. If we d o not have a closed form for the goal d istribution we are not able to d raw sam ples from it. H owever, we can d raw sam ples from another d istribution, for instance the uniform , shown in picture b). P icture c) shows the sam ples, weighted accord ing to the Im portance S am pling. We com pute the probability m ass falling in an interval p by sum m ing the weights of the sam ples falling in the interval. If using a high num ber of sam ples the probability m ass com puted from a) and c) are sim ilar, becom ing eq ual as the num ber of sam ples goes to in nity. O bserve that, in case we are able to d raw sam ples from the target d istribution, (i) (i) such that p(x0 :t | z1:t ) (x0 :t | z1:t ) then all of the weights are the sam e, and the variance of w (i) is 0 . A n intuitive ex planation of how the im portance sam pling principle work s is given in Figure 2 .2 . S eq u ential Im p o rtanc e S am p ling B y restricting to the set of M ark ovian system s, and in particular focusing the choice on a particular class of im portance functions, such that (x0 :t | z1:t ) = (xt | x0 :t1 , z1:t ) (x0 :t1 | z1:t1 ) (2 .7 )

it is possible to recursively com puting the im portance weights, without revising the past generated trajectories, since
(i) wt

p(x0 :t | z1:t ) (x0 :t | z1:t ) p(zt | xt )p(xt | xt1 ) p(x0 :t1 | z1:t1 ) (xt | x0 :t1 , z1:t )
(i) (i) (i) (i) (xt (i) (i) (i) (i) (i) (i) (i)

(i)

(x0 :t1 | z1:t1 ) wt1


(i)

(i)

(i)

p(zt | xt )p(xt | xt1 ) |


(i) x0 :t1 , z1:t )

(2 .8 )

Where = 1/ p(zt | x0 :t1 , z1:t1 ) is a norm aliz ation factor. S everal approaches select the im portance function to be the transition m od el p(x t |xt1 ). A ccord -

Bayesian S tate E stim atio n


(i)

20

ing to the im po rtanc e sam pling princ iple the weights w t c an be c o m pu ted as fo llo ws: wt
(i)

p(xt | z1:t ) p(xt | xt1 )


(i) (i) (i)

(i)

=
(i) (i) (i)

p(zt | xt )p(xt | xt1 )p(xt1 | z1:t1 ) p(xt | xt1 )


(i) (i) (i)

wt1 p(zt | xt ).

(i)

(2 .9 )

S am p lin g Im p o rtan c e R esam p lin g T he d irec t u se o f a S eq u ential Im po rtanc e S am pling lter req u ires a hu ge nu m ber o f sam ples, sinc e as the system evo lves all o f the partic les bu t o ne will have a high weight. Fo r this reaso n the S am pling Im po rtanc e Resam pling (S IR) lter [2 1 ] has been intro d u c ed . A S IR lter, seq u entially pro c esses the o bservatio ns z t and the pro prio c eptive sensings ut as they are available, by u pd ating a set o f sam ples representing the estim ated d istribu tio n p(x0 :t | z1:t , u0 :t ). T his is d o ne by perfo rm ing the fo llo wing three steps: 1 . Sampling: T he nex t generatio n o f partic les {x t } is o btained by the (i) previo u s generatio n {xt1 }, by sam pling fro m a pro po sal d istribu tio n (xt | x0 :t1 , z1:t , u0 :t ). 2 . Impo rtance W eigh ting: An ind ivid u al im po rtanc e weight w (i) is assigned to each partic le, ac c o rd ing to the IS princ iple w
(i) (i) (i)

(i) (i) p(xt wt1 (i) (xt

| x0 :t1 , z1:t ) | x0 :t1 , z1:t )


(i)

(i)

(2 .1 0 )

T he weights w (i) ac c o u nt fo r the fac t that the pro po sal d istribu tio n in general is no t eq u al to the tru e d istribu tio n o f su c c esso r states. 3 . Resampling: Partic les with a lo w im po rtanc e weight w are typic ally replac ed by sam ples with a high weight. T his step is nec essary sinc e o nly a nite nu m ber o f partic les are u sed to appro x im ate a c o ntinu o u s d istribu tio n. Fu rtherm o re, resam pling allo ws to apply a partic le lter in situ atio ns in which the tru e d istribu tio n d i ers fro m the pro po sal o ne. Please no te that a S IR lter here d esc ribed , assu m es the pro po sal to be su itable fo r seq u ential estim atio n. T his m eans that () satis es E q . 2 .7 . In princ iple, o ne wo u ld evalu ate po int wise and sam ple fro m the target d istribu tio n p(xt | z1:t , u0 :t ), bu t this is hard , d u e to the fo llo wing reaso ns:

Bayesian S tate E stim atio n it is u su ally not available in closed form and , it d epend s on the whole inpu t history z1:t and u0 :t , u p to time t.

21

In the nex t of this section, we d iscu ss a cru cial problem a ecting the particle lters: particle d epletion. S u ch a problem can be d etected by the analysis of some ind icators of the lter behavior, while it can be bou nd ed by a proper choice of the proposal d istribu tion. Partic le D ep letio n W hile the resampling step is need ed for concentrating the compu tational e ort of the lter in state space reg ions having a hig h lik elihood , it introd u ces ad d itional problems. T his problem becomes evid ent when u sing an u ninformed proposal d istribu tion, while the observations are a ected by a small noise. M ost of the samples g enerated in the u ninformed sampling step will be reward ed with a low weig ht in the importance weig hting phase, and they are lik ely to be su ppressed by resampling . In some d eg enerated situ ation, after the resampling step only one particle is retained . T his behavior can lead the lter to converg e to the wrong solu tion, since the mu ltimod al posterior is not ad eq u ately covered by the samples. T his problem is k nown of as particle d epletion [7 0 ]. T wo are the techniq u es for lessening particle d epletion: sampling from a proposal which is closer to the targ et d istribu tion, and ad d ing arti cial noise to the observation mod el. T he rst solu tion is stru ctu ral, since the choice of a better proposal d istribu tion mak es the importance weig hts valu e to be similar for all of the particles, in fact limiting the particle su ppression in the resampling stag e. T he second solu tion can lead to a work ing lter, bu t the introd u ction of arti cial noise d ecreases the accu racy of the estimate with respect to the one achievable by u sing the orig inal observation mod el. M oreover, when u sing a better proposal d istribu tion particle d epletion can be lessened by red u cing the nu mber of resampling actions. S ome alternated ltering schemes have been proposed , in ord er to lessen particle d epletion, lik e the au x iliary particle lter B y P itt and S hephard [5 4 ], however the improvements achievable u sing these techniq u es are orthog onal to the selection of an improved proposal. N u m ber o f E ec tiv e S am p les L iu [3 9 ] introd u ced the so-called e ective nu mber of particles Ne to estimate how well the cu rrent particle set represents the tru e posterior. T his q u antity is compu ted as Ne = 1
N i= 1

w(i)

2.

(2 .11)

T he intu ition behind Ne is as follows. If the samples were d rawn from the tru e posterior, the importance weig hts of the samples wou ld be eq u al to each other,

Bayesian S tate E stim atio n

22

due to the im portanc e sam pling princ iple. T he worse is the approx im ation the hig her is the varianc e of the im portanc e weig hts. S inc e N e c an be reg arded as a m easure of the dispersion of the im portanc e weig hts, it is a useful m easure to evaluate how well the partic le set approx im ates the true posterior. C h o ic e o f th e o p tim al p ro p o sal d istrib u tio n im portanc e func tion has been introduc ed in [3 9 ]: (xt |z1:t ) = p(xt |xt1 , zt ).
(i)

T he optim al seq uential

(2.1 2)

T he optim ality has to be intended as m inim iz ing the varianc e of the im portanc e weig hts. If drawing from the optim al proposal distribution, the im portanc e weig ht w(i) for each partic le i is c om puted ac c ording to E q . (2.1 0 ):
(i) wt

= =

p(x0 :t | z1:t ) (x0 :t | z1:t )


(i) (i) (i) (i)

(i)

= = [using E q .2.8 ]
(i) (i)

(2.1 3 )

p(x0 :t | z1:t ) p(xt | xt1 , zt )


(i)
(i) (i)

p(zt | xt )p(xt | xt1 )


p(zt |xt )p(xt |xt1 )
(i)

(i)

wt1

p(zt | xt )p(xt |xt1 )d xt


(i) p(zt |xt1 )

(i)

= wt1 p(zt | xt1 ).

(i)

(i)

(2.1 4 )

R ed u c in g th e S am p lin g S p ac e D im en sio n D ue to the ex tensive representation of the probability distribution used by the partic le lters, the num ber of partic les req uired for a g ood approx im ation of a ltering distribution, g rows ex ponentially with the dim ension of the state to estim ate. In this c ase, however, it is possible to address the problem by Rao-B lack welliz ation [1 1 ]. T his techniq ue c an be applied when the dynam ic B ayesian network of the system has a partic ular struc ture [1 2]. It c onsists in partitioning the state spac e X in two subspac es X a and X b . T he partition has to be m ade ac c ording to the struc ture of the system , ensuring that, g iven a state x t X , its projec tion a xa t X depends only on the previous state x t1 , the c urrent c om m and ut b and the c urrent observation zt . T he x projec tion xb t X should be updated a analytic ally and e c iently onc e xt is k nown. T his allows to sam ple only from X a , in fac t dec reasing the e ec tive sam pling spac e dim ension.

You might also like