A Tutorial on Particle Filters for On-line Nonlinear /Non-Gaussian Bayesian Tracking * Simon Maskell Qinetiq Ltd and Cambridge University Engineering Department, UK. s.maskell@signal.qinetiq.com Neil Gordon QinetiQ Ltd, UK. n.gordon@signal.qinetiq.com September 2001 1 Introduction Bayesian methods provide a rigorous general framework for dynamic state estimation problems. The Bayesian approach isto construct the paf of the state vector based on all avalable informatio. This pdf summaries the current state of knowledge about tbe stave vector and from it the optimal (with respect to whatever cos funtion the ser chooses) course of action can be determined. For the linear/Gaussian ‘etimation problem, the required pdf remains Gaussian at every iteration of the filter, and the Kalman Slvr relations propagate and update the mean and covariance ofthe distribution. For a noninesr/ non- Gaussian problem there is in general no analytic (closed form) expression forthe requited paf. However, for many applications these modelling assumptions are strongly implied by considerations of realism For exarnple, bearings only tracking in a Cartesian co-ordinate system implies a nonlinear relationship between the’ state and the measurement. Similarly, if measurement sensors produce occasional gross cevors (eg radar gat), a heavy tailed non-Gaussian error model could be appropriate. In the following notes we begin with a description ofthe nonlinear /aon-Caustian tracking problem and its optimal Bayesian solution. Since the optimal solution i intractable, several different approximation strategies are then described. These approaches include the extended Kalman filter and particle iter. ‘These notes are ofa tutorial nature and t,o facilitate easy implementation, ‘pseudo-code” for algorithms has been included at relevant points. 2 Nonlinear Bayesian Tracking ‘To define the problem of tracking, consider the evolution of the state sequence {x,,k € N} of a target, given by xe = falas ¥4-2) o where fy RY xR" + RY is a possibly nonlinear function of the state x1, {vith €N} is an Lid process noise sequence, nz;ne are dimensions of the state and process noise vectors, respectively and N {s the set of natural numbers. The objective of tracking isto recursively estimate x, from measurements, m= by(xyms), 2) ~@Coprigh Gineiq Lid 2001, Published withthe permson of QinetiQ Lid. This paper is condensed version of 2nwhere by :% x R% + Ris a possibly nonlinear function, {my, € N} 's an iJ.d measurement noise sequence, and n,,ftq are dimensions of the measurement and measurement noise vectors, respectively. In particular, we seek filtered estimates of x, based on the set of all available measurements 214 = {2:, Tyeeey} up to time k. From a Bayesian perspective, the tracking problem is to recursively calculate some degree of belief in the sate xp at time f, taking different values, given the data zy... Thus, it is required to construct the pdf p(xaltin). It is assumed that the intial paf, p(xol2o) = plo), of the state vector, also known as the prior, is available (zo being the set of no measurements). Then, in principle, the pdf p(xslzis) may be obtained recursively in two stages: prediction and update. Suppose that the required pdf p(x,-1|21.4-1) at time k~1 is available. The prodiction stage involves using the system model (1) to obtain the prior pdf of the state at time k via the Chapman-Kolmogorov equation: seuss) = f sbeebrea)e Geis fsa-s ei ® At time step k, a measurement 25 becomes available, and this may be used to update the prior (update stage) via Bayes’ rule: plsaleis) = @ where the normalising constant esses) =f ranbse)efesa-s)eee © depends on the lkeibood function p(sa[xe). In the update stage (4), the measurement 2p is used to ‘modify the prior density to obtain the required posterior density ofthe curent state. ‘The recurrence relations (3) and (4) form the basis for the optimal Bayesian solution. This recursive propagation of the posterior density is only a conceptual solution in that in genera, it cannot be de- ‘termined analytically. Solutions do exist in a restrictive set of cases, including the Kalman filter. We now describe how, when the analytic solution is intractable, extended Kalman filters and partile Stars approximate the optimal Bayesian solution 3 Extended Kalman Filter ‘The most well known, and certainly the most popalar, approach to implementing recursive nonlinear filters is the extended Kalman filter (EKF), see [19]. This approach is applicable to nonlinear models with additive Gaussian noise. It sa linearisation technique based ona first order Taylor series expansion of the nonlinear system and measurement functions about the current estimate ofthe state. This requires ‘that both fy and hy be differentiable fnetions oftheir vector arguments. Pioxpeiftiaca) & N(Xe-aimiaaji-t,Feosihos) © Peltie-1) & N(KesMajea, Pant) O) = Neus fam apna)s Onna + Fe Phaae a FP) ® psalm) A M84; Pain) ~ KPa) ®) N (xa; mga + Ke(te = Ba (maya), Pa where f,() and hy() are nonineas functions and A and 2 are local inavsations ofthese noaliear Fanctions (i. masrices): a9) 22ay (a2) k= a3) ‘The EKF as described above utilises the first term in a Taylor expansion of the nonlinear function. A higher order EKF that retains further terms in the Taylor expansion exists, but the additional complexity hhas probibited its widespread use. Recently, the unscented transform has been used in an EKF framework [20, 31, 32). The resulting filter, known as “Unscented Kalman Filter”, considers a set of points that are deterministically selected from ‘the Gaussian approximation to p(xslti.s). These points are all propagated through the true nonlinearity ‘and the parameters of the Gaussian approximation are then reestimated. For some problems, this ‘ter has been shown to give better performance than a standard EKF since it better approximates the nonlinearity; the parameters of the Gaussian approximation are improved. However, the EKF always approximates p(xu[zi) to be Gaussian. Ifthe true density is non-Gaussian (eg, if is bimodal oF heavily skewed) then a Gaussian can never describe it well, In such cases, approximate grid-based fers and particle Sere wil yield an improvement in performance in comparison to that of an EKF (2 4 Particle Filtering Methods 4.1 The Sequential Importance Sampling (SIS) Algorithm ‘Tae Sequential Importance Sampling (SIS) algorithm is a Monte Carlo (MC) method that forms tbe basis for most sequential Monte Carlo iters developed over the past decades - see (13, 14) This sequen- tial Monte Carlo (SMC) approach is known variously as bootstrap filtering [17], the condensation algo- rithm (25), particle Sitering (6), ineeracting particle appreximations (10, 11) and survival of the Gittest (21) Iisa technique for implementing a recursive Bayesian fier by Monte Carlo simulations. The key dea s to represent the required posterior density function by a set of random samples with associated weights and to compute estimates based on these samples and weights. As the number of samples becomes very large, this Monte Carlo characterisation becomes an equivalent representation to the usual functional description ofthe posterior pdf, and the SIS flter approaches the optimal Bayesian estimate. In order to develop the details ofthe algorithm, let {xh,,wh)i% denote a Random Measure that charac teriss che posterior pdf p(xoaltia), where (xi4,f = 0,-- Ns} 18a set of support points with associated weights {uf,i=1,...,N,) and xox = {x,j = 0,...,k} isthe st of all states up to time k. The weights are normalised such that’, wf = 1. Thea, the posterior density at can be approximated as Plsoaltia) © So whslxos ~ Xba) a) ‘We therefore have a discrete weighted approximation to the true posterior, p(xoa[ei). The weights are chosen using the principe of Importance Sampling [3,12]. Ths principe relies on the following: Suppose ‘p(2) « r(2) isa probability density from which iis dificult to draw samples, but for which x(z) can be evaluated (and so p(2) up to proportionality). Also, lt 24 ~ g(2), #= 1y..-y.Ny be samples that are easly generated from a proposal (-), called an Importance denaty. ‘Then, a weighted approximation to ‘he density n() is given by He) 2 Soulsee~ 25) as) where 3 wal as)is the normalised weight of the i** partile. So, # the samples, x),,, were drawn from an importance density, g(xo.4/zi-.) then the weights in (14) are defined by (16) to be ef oc POMpaltus) an aGealzoa) Returning tothe sequential case, at each iteration one could have samples constituting an approximation to p(xox-1 214-1), and want to approximate p(Xoa|ti:e) with a new set of samples. If the importance density is chosen to factorise sch that axoala = afsabxo-1.14)4(on-aleia-2) as) ‘then one can obtain samples x5, ~ 9(Xou/2i) by augmenting each of the existing samples xj,4_; ~ a(von-rleia-1) with the new state x, ~ a(xa[son-i,2ia). To derive the weight update equation, ‘P(Rok|2i-4) is first expressed in terms of p(xo.e—1|21:4=1), P(ze|xe) and p(xeber-1)- x PlEelKou, Z4—1) P (Xo |zree—1) a Pris) = Plzabton, tis) eabtosess 2-1)» Coast cy = Peete alrite alts) & p(ealss) p(xsbea-1) (Haase) a) By substituting (18) and (19) into (17), the weight update equation can then be shown to be fc PUtabeh) p iteh a) POSa-aleia—a) aC 8) —) Pleslxp)eGeh eh) wh elisa) (20) re ‘Furthermore, if ¢(xx}¥o4-1,23:4) = q(xa[s—1,24), then the importance density becomes only dependent ‘on the x41 and z. This is particularly useful in the common case when only a Sltered estimate of ‘p(xsl2ie) is required at each time step. From this point on, we shall assume such a case, except when ‘explicitly stated otherwise. In such scenarios, only x}, need be stored, and so one can discard the path, baci and history of observations, 2-1. The moditied weight is then see uf, Beason aye) and the posterior filtered density p(xs|21.x) can be approximated as x lxalzran) © Do whdloe - xi) (22) ‘where the weights are defined in (21). It can be shown that as Ny -> co the approximation (22) approaches the true posterior density p(xplzi.1). ‘The SIS algorithm thus consists of recursive propagation of the weights and support points as each ‘measurement is received sequentially. A pseudo-code description ofthis algorithm is given by algorithm 1. 214ALGORITHM 1: SIS PARTICLE FILTER txi.wit; * FORi= = SIS [{xh1, 4h» = Draw xf ~ a(elh_s.t) — Assign the particle a weight, w{, according to (21) * END FOR 4.1.1 Degeneracy Problem ‘A common problem withthe SIS particle filter isthe degeneracy phenomenon, where after afew iterations, all but one particle will have negligible weight. It has been shown [12] that the variance of the importance ‘eights can only increase ove time, and thus tis impossible to avoid the degeneracy phenomenon. Tis degeneracy implies that a large computational effort is devoted to updating particles whose contribution ‘to the approximation to p(xs[21.4) is almost zero. A suitable measure of degeneracy of the algorithm is the efecive sample size Nery introduced in (3) and [24 and defined at x, Nu aay (23) where wy! = p(xilzs)/aGxkbxt1,21) is referred to as the “true weight”. ‘This cannot be evaluated exactly, but an estimate Nazy of Nes can be obtained by on M= ae @ 0 SEP ‘where wf is the normalised weight obtained using (20). Notice that Nery < Ny, and small N,yy indicates levere degeneracy. Cleary, the degeneracy problem isan undesirable effect im particle ter. ‘The brave {ree approach to reducing ite efecto use a very large Ni This soften impractical, and so we rely on two other methods: a) Good choice of importance density, aad b) Use of resampling. These are described nex. 4.1.2 Good Choice of Importance Density ‘The first method involves choosing the importance density, @(xs)x}_1,4), to minimise Var(w') so that [Ney is maximised. The optimal importance density function which minimises the variance of the true ‘weights, wf, conditioned upon xj_, and 24 has been shown [12] to be axebeL a talone = plcelhasts) _ PlGalce x = Substietion of (25) at (21) yields wh iP (eli. = wha f rleateh) (nhs) be 2s) 2S‘This choice of importance density is optimal ince for a given .., tf takes the same value whatever sample is drawn from q(xalx{_,,4)op- Hence, conditional on x}_,, Var(wf*) = 0. This is the variance of the diferent w resulting from different sampled xf, ‘This optimal importance density suffers from two major drawbacks. It requires the ability to sample ‘rom p(xelx{_y,4) and to evaluate the integral over the new state. In the general case it may not be straightforward to do either of these things. There are two cases when use of the optimal importance density is possible. ‘The first case is when xy is a member of a finite set. In such cases, the integral in (25) becomes a sum ‘and sampling from p(s[x}_1,21) is possible. An example of an application when xy is a member of a finite set isa Jump-Markov Linear System for tracking maneuvering targets (15). Analytic evaluation is possible for a second class of models for which p(xalx{_ 1,21) is Gaussian (12, 9 ‘This can oocur Ifthe dynamics are nonlinear and the measurements linea. Such a system's given by Xe = filXea) + Vea, Venn ~ N(¥a-15 00,21, Qe-1) m= Atm, Be ~ MlaiOnnis Fs), . ‘where fy : 3% —+ > is a nonlinear function, Hy € R"+%" is an observation matrix, and vii and my are mutually independent iid Gaussian sequences with Qs. > 0 and Ry > 0. Defining By = Oy, + HER Ae (28) mg = De (Qj! falas) + AP RG ee)» (29) one obtains plosalnaa te) = M(xasmms, Ei) (30) and PCzeLXe-1) = N (en; Hatin a1), Oana + HaReH). (1) For many other models, such analytic evaluations are not possible. However, it is possible to construct suboptimal approximations to the optimal importance density by using local linearisation techniques [12] Such linearisations use an importance density that is a Gaussian approximation to p(xe[xs112s). A~ ‘other approach is to estimate a Gaussian approximation to p(e[xe-1,24) using the unscented trans- {orm (90). ‘The authors’ opinion is that the additional computational cost of using such an importance density is often more than offset by a reduction in the number of samples required to achieve a certain Jevel of performance. Finally, itis often convenient to choose the importance density to be the prior. g(a 1528) = pala) (82) Substitution of ($2) into (21) then yields ‘wh tsa). (33) ‘This would seem to be the most common choice of importance density since it is intuitive and simple to implement. However, there are a plethora of other densities that can be used and the choice isthe crucial design step for a particle filter. 4.1.3 Resampling ‘The second method by which the effects of degeneracy can be reduced is to use resampling whenever ‘a significant degeneracy is observed (le. when Ney, falls below some threshold, Nr). The basic idea of resampling is to eliminate particles which have sinall weights and to concentrate on particles with large 216weights. The resampling step involves generating a new set (x{"}!% by resampling (with replacement) AN, times from an approximate discrete representation of p(xultua) given by P(xelaiia) & S> wh8 xe — x}) (34) s0 that Pr(x{" = x}) = wh. The resulting sample isin fact an iid sample from the diserete density (34), and so the weights are now reset to w{ = 1/N,. It is possible to implement this resampling procedure ia OWN.) operations by sampling N, ordered uniforms using an algorithm based on order statistics (29,6 Note that other efficent (in terms of reduced MC variation) resampling schemes such as stratified sampling and residual sazpling [24] may be applied as alternatives to this algorithm. Systematic resampling [22] is the scheme preferred by the authors (since iti simple to implement, takes O(N) time and minimises ‘the MC variation) and its operation is described in algorthen 2, where Ufa, 6] s the Uniform distribution ‘on the interval (0,5. For each resampled particle x", this resampling algorithm also stores the index of its parent, denoted by #. "This may appear unnecessary here and is, but it proves useful in section 4.2.2 ‘A generic particle filter is then as described by algorithm 3, ALGORITHM 2: RESAMPLING ALGORITHM [Gd wi, #)fS,) = RESAMPLE (x, 6} « Tntialise the CDF: = 0 + FORI=2:4, ~ Construct CDF: 6 = ¢i-3 +h + END FOR + Start at the bottom of the CDF: Draw a starting point: w ~ U0, + FORJ=1:™, ~ Move along the CDF: = WHILE u; > 4 a ieiel ~ END WHILE — Assign sample: xj" ut NG=1)ALGORITHM 3: GENERIC PARTICLE FILTER oh wid PE (fama) « FORi=1:N, ~ Draw xf ~ a(xabeh ast) — Assign the particle a weight, w{, according to (21) END FOR © Calculate total weight: t FORi=1:%, ~ Normalise: wf = END FOR © Calculate N77 using (24) — Resample using algorithm 2: [{x},.},—}h] = RESAMPLE ({x},wi} 4) ENDIF ‘Although the resampling step reduces the fics of the degeneracy problem, it introduoes other practical ‘problems. Fist t limits the opportunity to paralelize since all the particles must be combined. Second, ‘the particles which have high weights w{ are statistically selected many times. ‘This leads to a loss of diversity among the particles ar the resltant sample will contain many repeated points. This problem, \nown as sample impoverishment, is geverein the case of small proces oise. Infact, for Ube cate of ery ‘small process noise all particles will collapse to a single point within a few iterations! Thirdly, since the diversity of the paths of the particles is reduced, any smoothed estimates based on the particles’ paths degenerate?. Schemes exist to counteract this effect. One approach considers the states for the particles to be pre-determined by the forward filter and then obtains the smoothed estimates by re-calculating the pacticles’ weights via a recursion from the final to the fist time step [16]. Another approach is to use MCMC [3 ‘There have been some systematic techniques proposed recently to solve the problem of sample impoverish- ‘ment. One such technique is the resample-move algorithm [18], Although this technique draws conceptu- ally on the same technologies of importance sampling-resampling and MCMC sampling, it avoids sample ‘impoverishment. It does this in a rigorous manner that ensures the particles asymptotically approximate samples from the posterior and s0 is the method of choice of the authors. An alternative solution to the same problem is regularisation [26], which is discussed in Section 4.2.3. This approach is frequently found to improve performance despite a les rigorous derivation and is included here in preference to the resample-move algorithm since its use is so widespread. = partie Siter ie ot entirely appropriate. Particle Biteving is a method well sited tothe eetimation of dysamie mates, If atie ates, which can be regarded as paraoe:rs, need tobe estimated then Slteroative approsches are poesary [7 23) Since the parca actually represent paths through the state space, by toring the trajectory taken by each particle, smoothed erimates ofthe state ean be obtalned [4 2184.1.4 Techniques for Circumventing the Use of a Sub-Optimal Importance Density eis often the case that a good importance density is not available. For example, ifthe prior (xe[xe—1) is used asthe importance density and is a much broader distribution than the likehood, p(zalkx), then only a few particles will have a high weight. Methods exist for encouraging the particles to be in the right place; the use of bridging densities (8) and progressive correction {27] both introduce intermediate distributions between the prior and likelitood. ‘The particles are thea re-weighted according to these intermediate distributions and resampled. This “herds" the particles into the right part of the state space. Another approach known as partitioned sampling [25] is useful if the likelihood is very peaked, but ean be factorised into a number of broader distributions. Typically, this occurs because each of the partitioned distributions are functions of some (not all) of the states. By treating each of these partitioned distributions in turn and resampling on the basis of each such partitioned distribution, the particles are ‘again herded towards the peaked likelihood. 4.2 Other Related Particle Filters “The Sequential Importance Sampling algorithm presented in Section V-A forms the bass for most particle filters that have been developed s fa. The various versions of particle filters proposed in the literature can be regarded as epecial case of this general SIS algorithm, ‘These special cases can be derived from the SIS algoritam by an approprace choice of importance sampling density and/or modification ofthe resampling ‘step. Below, we present three particle filters proposed in the literature and show how these may be derived from the SIS algorithm. The particle filters considered are (i) Sampling Importance Resampling (SIR) filter (ii) Auxiliary Sampling Importance Resampling (ASIR) filter, and (iii) Regularised Particle filter (PF) 4.2.1 Sampling Importance Resampling Filter ‘The Sampling Importance Resampling (SIR) filter proposed in [17 i a Monte Carlo method that can be applied to recursive Bayesian filtering problems. The assumptions required to use the SIR filter are very ‘weak. The state dynamics and measurement functions, fy(,-) and y(,-) fn (1) and (2) respectively, need to be known, itis required to be able to sample realisations from the process noise distribution of v1 ‘and from the prior. Finally, the likelihood function p(zy[x) needs to be aveilabe for pointwise evaluation (et least up to proportionality). The SIR algorithm can be easily derived from the SIS algorithm by an appropriate choice of (i) The importance density: q(xa{xb_.,21) is chosen to be the prior density ‘p(xelx}_,), and (i) Resampling step: to be applied at every time index. ‘The above choice of importance density ienplies that we need samples from p(xslxi..)- A sample x} ~ cae) can be generated by fit geatng« roc min spl vj ~ Fina) aad eng xf = FeG_1.¥h 1), where po() is the pdf of vp. For this particular choice of importance density, it Isevident that the weights are given by sof cc wha poe a 3) However, noting that resampling is applied at every time index, we have w{_, =1/N Vi and so sof oc nual). (36) ‘The weights given by the proportionality in (36) are normalised before the resampling stage. An iteration ‘of the algorithm is then described by algorithm 4.ALGORITHM 4: SIR PARTICLE FILTER Ext teh 2] = SIR Ht © FORI= Ny ~ Draw x ~ p(eeb ~ Calculate wi = plauixh) + END FOR ‘Calculate total weight: ¢ = SUM {uf} + FORE=1:%, Normalise: wf = tf + END FOR ++ Resample using algorithm 2: [{x},w}, ~}0] = RESAMPLE [{x}, is As the importance sampling density for the SIR filter is independent of measurement 2p, the state space is explored without any knowledge of the observations. ‘Therefore, this filter can be ineficient and is sensitive to outliers. Furthermore, as resampling is applied as every iteration, this can result in rapid loss of diversity in particles. However, the SIR method does have the advantage that the importance weights are easily evaluated and the importance density can be easily sampled. 4.2.2 Auxillary Sampling Importance Resanopling filter The Aulary Sampling Importance Resampling (ASIR) flter was introduced by Pitt & Shephard [28] ‘8 variant ofthe standard SIR Slr. This flter can be derived from the SIS framework by introducing ‘an importance density g(xasei.a) which samples the pair (x1,47}%:, where? refers to the index of the particle at F~1. By applying Bayes rule, a proportionality can be derived for p(x il Peayflta) & plealxe)p(ae,iltia1) = plsalr)eCsalixi:a-1)pCiln-2) lex) PCP )0h a er) ‘The ASIR filter operates by obtaining a sample from the joint density p(xiy if), and then omitting the indies i in the pair (x4,4) to produce a sample (x), fom the marginalised density p(xalais)- ‘The importance density used to draw the sample {x}, i'}{¥, is defined to satisfy the proportionality aCe ifn) x plsalish)oCxalh 04-2 (38) ‘where 4, is some characterisation of xa, given xj... ‘This could be the mean, in which case p= Eixe lx} or a sample, ah ~ poxelsh_,)- By writing ay iT2ia) = ailera)eCaeli zis 9) 2/10and defining alativars) £ vlads a), (40) it follows fom (38) hat alien) o Pea nok 1 (ay ‘The sample, (xi), is then asigned a weight proportions! tothe ratio ofthe RUS of (37) to (38): wo wf Peale at) . arts) pieabet) ~ pleat) te) ‘The algorithm then becomes that described by algorithan 5, ALGORITHM 5: AUXILIARY PARTICLE FILTER Uexhseb hia) = APF [ted + FORE=1:N% ~ Catelace of ~ Calelate wf + END FOR eae) oe weal u * Calculate total weight: t = SUM [{w}}?%] + FORI=1:N, ~ Normale: wf = ¢-tu + END FOR + Resample using algorithm 2: [{—, = FORJ=1:N, ~ Draw xj ~ o(xuli, aux) = p(xalx{_;), a8 in the SIR filter. ~ Assign weight wf sing (42) = END FOR * Calculate total weight: ¢ = SUM ({wi}2%] + FORI=1:4, ‘= RESAMPLE [ix{,wi)!% ~~ Normalise: wi = tte, © BND FOR, Although unnecessary, the original ASI fr as proposed in [28] consisted of one more sep, namely a resampling sag, to produce an id sample {xf}, mith equal weg, 2aCompared to the SIR fiter, the advantage of the ASTR Slter is that i: naturally generates points from the sample at k— 1 which, conditioned on the current measurement, are most liktly to be close to the tre state. ASIR can be viewed as resampling atthe previous time step, based on some point estimates, ab that caracterie poealee). I the process noe is small, 0 peas) 8 wel characterised by pl, ‘then ASIR is often not so sensitive to outliers as SIR, and the weights w{ are more even, However, if the process noise is large, a single point does not characterise p(xu}xi...) well and ASIR resamples based on ' poor approximation of p(xs[x}). In such scenarios, the use of ASIR then degrades performance. 4.2.8 The Regularised Particle Filter Recall that resampling was suggested in section 42.1, as a method to reduoe the degeneracy problem ‘hich is prevalent in particle filters. However, it was pointed out that resampling in turn introduced ‘other problems, and in particular, the problem of los of diversity among the particles. This arises due +o the fact that in the resampling stage, samples are drawn from a discrete distribution rather than a continuous one. If this problem isnot addressed propery, it may lead to “particle collapse”, where all Ny particles occupy the same point in the state space, giving a poor representation ofthe posterior density. ‘A modified particle filter known as the Reguarised Particle Filter (RPF) was proposed [26] as a potential ‘solution to the above problem. ‘The Regularised Particle filter is identical to the SIR filter except for the resampling stage. The RPF resamples from a continuous approximation of the posterior density p(xlz)..) while the SIR resamples ‘from the discrete approximation (34). Specifically, in the RPF, samples are dravn from the approximation outta) = SS ubKvou —x6) as) where Ku) = EK (F) “) is the rescaled Kernel density K(), h > 0 is the Kernel bandwidth (a scalar parameter), ng is the dimension of the state vector x, and uj, ¢= 1,...,Ny, are normalised weights. The Kemel density ie a symmetric probability density function such that [xk eac=0, | tniPK (we <00 ‘The Kernel K() and bandwidth h are chosen to minimise the Mean Integrated Square Error (MISE) between the true posterior density and the corresponding regularised empirical representation in (43), defined at mase(e) = Bf fasta) —ptsaara? aa] as) where £(-|-) denotes the approximation to p(xs)21.) given by the RHS of (43). In the special case of an ually weighted sarople, the optimal choice of the Kerel ete Epanecnikov Kere! 26, 20.) it) <2 km={ 5 4) where ¢, is the volume of the unit hypersphere in 8%. Furthermore, when the underlying density is Gaussian with a unit covariance matrix, the optimal choice for the bandwidth is [26) hope = ANT an) A= [Beg} (ng + 4)(2V A)" ] (48) ‘Though the sesults of (46) and (47)-(48) are optimal only in the special case of equally weighted particles and underlying Gaussian density, these results can still be used in the general case to obtain a suboptimal 2n2filter. One iteration ofthe RPF is described by algorithm 6. The RPF only difers from the generic particle ‘ter described by algorithm 3 as a reeult of the addition of the regularisation steps when conducting the resampling. Note also that the calculation of the empirical covariance matrix Sj, is cartied out prior to the resampling and so is a function of both the x} and wf. This is done since the accuracy of any estimate of a function of the distribution can only decrease as a result of the resampling - if quantities such as the mean and covariance of the samples ate to be output then these should be calculated prior to resampling. ALGORITHM 6: REGULARISED PARTICLE FILTER (ei ia) RP (xh st0f 274] + FORI=1:%, ~ Draw x} ~ a(euh a) ~ Assign the particle a weight, wi, according to (21) + END FOR Calculate total weight: # = SUM [{w{}% + FORI=1:%, ~ Normalise: wf = ttuf + END FOR Calculate N77 using (24) oN < Nr ~ Calculate the empitical covariance matrix Sof (xj, wf}! ~ Compute Dy such that DxDJ = Sy. — Resample using algorithm 2: [{x},w,~}%] = FORI=1:%, + Draw e' ~ K from the Epanechnikov Kernel + xT = xh + AopeDae! ~ END FOR o END RESAMPLE [(xj, wi} By following the above procedure, we generate aa iid random sample {xj"}, drawn from (48). 1 terms of complerty, the RPF is comparable to SIR since it only requires N, additional generations from the Kernel K() at each time step. The RPF has the theoretic disadvantage thatthe samples are no longer guaranteed to asymptotically approximate those from the posterior. In practical scenarios, the PF performance is better than the SIR in cases where sample impoverishment is severe, for example ‘when the process noise is small 2n35 Conclusions For a particular problem, if the assumptions of the Kalman filter hold then no other algorithm can out-perform it. However, in a variety of real scenarios, the assumptions do not hold and approximate ‘techniques must be employed. ‘The extended Kalman filter approximates the models used for the dynamics and measurement process, in order to be able to approximate the probability density by a Gaussian. Particle filtering approximates the density directly as a finite number of samples. A number of different types of particle flter exist ‘and some have been shown to outperform others when used for particular applications. However, when ‘designing a particle filter for a particular application, itis the choice of importance density that is critical References {i} § Arulampalam, $ R Maskell, N J Gordon & T Clapp, A Tutorial on Particle Filters for On-line ‘Nonlinear/Now-Gaussian Bayesian Tracking, To appear: IEEE Transaction on Signal Processing [2] S Arulampalam & B Ristic, Comparison ofthe Particle Filter with Range Parameterised and Modied Polar EKF’s for Angle-Only Tracking. Signal end Data Processing of Small Targets 2000, SPIE Volume 4048, pp.288-209. [3] N Bergman, Recursive Bayesian Estimation: Navigation and Tracking Applications. 1999, PhD The- sis, Linkoping University, Sweden. [4] N Bergman, A Doucet & N Gordon, Optimal Estimation and Cramer-Rao Bounds for Partial Non- Gaussian State Space Models. Ann. Inst. Statist. Math., 2001, Volume 83(1), pp 97-112. [5] BP Carlin, N G Polson & DS Stoffer, A Monte Carlo approach to nonnormal and nonlinear state- space modelling, Journal of the American Statistical Association, 1992, 87(418), 493-500. [6] J Carpenter, P Cliford & P Fearnhead, Improved Particle Filter for Nonlinear Problems, In IEE ‘Proceedings on Radar and Sonar Navigation, 1999. [7] T Clapp, Statistical Methods for the Processing of Communications Data, PhD thesis, Department of Engineering, University of Cambridge, 2000 (6) T Clapp & S Godsill, Improvement strategies for Monte Carlo particle ers. in Sequential Monte Gorlo Methods in Practice, eds A Doveet, J F G de Freitas ond N J Gordon, 2001, Springer-Verlag: ‘New York. [8] P Del Moral, Measure Valued Proceses and Interacting Particle Systems. Application to Nonlinear Filtering Problems. Annale of Applied Probability, 1996, Volume 8, No2, pp 438-495. [00] D Crisan, P Del Moral & TJ Lyoas, Nonlinear Filtering Uing Branching and Interacting Particle Systems. Markov Processes ond Related Fields, 1999, Volume 5, No 3, pp 208-319. (11] P Det Morai, Nonlinear Filtering: Interacting Particle Solution, Markov Proceses and Related Fields Volume 2, No 4, pp 555-580. [22] A Doucet, On Sequential Monte Carlo Methods for Bayesian Filtering. 1998, Technical Report, ‘University of Cambridge, UK, Department of Engineering. [13] A Doucet, J F G de Freitas & N J Gordon, An introduction to sequential Monte Carlo methods, in ‘Sequential Monte Carlo Methods in Practice, eds A Doucet, J F G de Freitas and N J Gordon, 2001, Springer-Verlag: New York. {14} A Doucet, $ Godsill & C Andriou, On sequential Monte Carlo sampling methods for Bayesian Altering, Statistics and Computing, 10(3), 197-208. ana[15] A Doucet, N Gordon & V Krishnamurthy, Particle Filters for State Estimation of Jump Markov {near Syatens, IEEE Transactions on Signal Procsing, March 2001, Volume 48 , No 3, pp 613 4 [6] $ Godsill, A Doucet & M West, Methodology for Monte Carlo Smoothing with Application to Time- ‘Varying Autoregressions. Prec. International Symposium on Frontiers of Time Series Modelling, 2000. [17 N Gordon, D Salmond & A F M Smith, Novel Approach to Nonlinear and Non-Gaussian Bayesian State Estimation. IEE Proceedings-F, 1993, Volume 140, pp.107-113. [18] W R Gilks & C Bereuini, Following a Moving Target - Monte Carlo inference for Dynamie Baysian ‘Models. Journal of the Royal Statistical Society, B, 2001, Volume 63, pp. 127-146. [19] A H Jazwinski, Stochastic Processes ond Filtering Theory. Academic Press: New York, 1970. (20) S Julier, A Skewed Approach to Filtering, Signal and Data Processing of Small Targets 1998, SPIE Volume 3375, pp.271-282. [Pt] K Kanazawa, D Koller & § J Russell Stochastic Simulation Algorithms for Dynamic Probabilistic ‘Networks. In Proceedings of the Eleventh Annual Conference on Uncertainty in AI (UAL '95), 1995, pp 346-351, [22] G Kitagawa, Monte Carlo Filter and Smoother for Non-Gaussian Nonlinear State Space Models. ‘Journal Of Computational and Graphical Statistics, 1996, Volume 5(1), pp.1-25. [23] J Liu & M West, Combined parameter and state estimation in simulation-based filtering, én Se- quential Monte Carlo Methods in Practice, eds A Doucet, J F G de Freitas and N J Gordon, 2001, Springer-Verlag: New York. 124) JS Liu& R Chen, Sequential Monte Carlo Methods for Dynamical Systems. Journal ofthe American ‘Statistical Association, 1998, Volume 98, pp.1032-1044, [25] 4 MacCormick & A Blake, A Probabilistic Exclusion Principle fr Tracking Multiple Objects. Proc “Int. Conf. Computer Vision, 1999, pp-572-578. (26) © Musso, N Oudjane & F LeGland, Improving Reguasised Pactile Filters. in: Sequential Monte Garlo Methods in Practice, eds A Doucet, JF G de Freitas and N J Gordon. 200, Springer-Verlag, New York. [2M] N Oudjane & C Musso, Progressive Correction for Regularized Particle Filters. Proc. Srd Interna ‘one Conference on Information Fusion, 2000. [28] M Pitt & Nv Shephard, Filtering via Simulation: Auxiliary Particle Filters. Journal of the American ‘Statistical Association, 1998, Volume 94(446), pp.590-599. (25) B Ripley, Stochastic Simulation, 1987, Wiley, New York. [30] R van der Merwe, A Doucet, JF G de Freitas &E Wan, The Unsoented Particle Filter. In Advances fn Neural Information Processing Systems (NIPS13), December 2000. {81} EA Wan & R Van der Merwe, The Unscented Kalman Filter for Nonlinear Estimation. In Proceedings ‘of Symposium 2001 om Adaptive Systems for Signal Processing, Communications and Control, Lake Louise, Alberta, Canada, Oct 2000, [s2] EA Wan & R Van der Merwe, To appear in Kalman Filtering ond Neural Networks, Chapter 7: ‘The Unscented Kalman Filter, Wiley Publishing, 2001. 2s