Convergence Analysis for the SMC-MeMBer and SMC-CBMeMBer Filters
Feng Lian,1Chen Li,2Chongzhao Han,1and Hui Chen1
Academic Editor: Qiang Ling
Received27 Feb 2012
Revised15 May 2012
Accepted22 May 2012
Published16 Aug 2012
Abstract
The convergence for the sequential Monte Carlo (SMC) implementations of the multitarget multi-Bernoulli (MeMBer) filter and cardinality-balanced MeMBer (CBMeMBer) filters is studied here. This paper proves that the SMC-MeMBer and SMC-CBMeMBer filters, respectively, converge to the true MeMBer and CBMeMBer filters in the mean-square sense and the corresponding bounds for the mean-square errors are given. The significance of this paper is in theory to present the convergence results of the SMC-MeMBer and SMC-CBMeMBer filters and the conditions under which the two filters satisfy mean-square convergence.
1. Introduction
Recently, the random finite-set- (RFS-) based multitarget tracking (MTT) approaches [1] have attracted extensive attention. Although theoretically solid, the RFS-based approaches usually involve intractable computations. By introducing the finite-set statistics (FISSTs) [2], Mahler developed the probability hypothesis density (PHD) [3], and cardinalized PHD (CPHD) [4] filters, which have been shown to be a computationally tractable alternative to full multitarget Bayes filters in the RFS framework. The sequential Monte Carlo (SMC) implementations for the PHD and CPHD filters were devised by Zajic and Mahler [5], Sidenbladh [6], and Vo et al. [7]. Vo et al. [8, 9] devised the Gaussian mixture (GM) implementation for the PHD and CPHD filters under the linear, Gaussian assumption on target dynamics, birth process, and sensor model. However, the SMC-PHD and SMC-CPHD approaches require clustering to extract state estimates from the particle population, which is expensive and unreliable [10, 11].
In 2007, Mahler proposed the multitarget multi-Bernoulli (MeMBer) [2] recursion, which is an approximation to the full multitarget Bayes recursion using multi-Bernoulli RFSs under low clutter density scenarios. In 2009, Vo et al. showed that the MeMBer filter overestimates the number of targets and proposed a cardinality-balanced MeMBer (CBMeMBer) filter [12] to reduce the cardinality bias. Then, the SMC and GM implementations for the MeMBer and CBMeMBer filters were, respectively, proposed in nonlinear and linear-Gaussian dynamic and measurement models. The key advantage of this approach is that the multi-Bernoulli representation allows reliable and inexpensive extraction of state estimates. The Monte Carlo simulations given by Vo et al. showed that the SMC-CBMeMBer filter outperforms the SMC-CPHD (and hence SMC-PHD) filter despite having smaller complexity under certain range of signal settings.
Although the convergence results for the SMC-PHD and GM-PHD filters were established by Clark and Bell [13] in 2006 and by Clark and Vo [14] in 2007, respectively, there have been no results showing the asymptotic convergence for the SMC-MeMBer and SMC-CBMeMBer filters. This paper demonstrates the mean-square convergence of the errors [15β17] for the two filters. In other words, given simple sufficient conditions, the approximation error of the multi-Bernoulli parameter set comprised of a set of weighted samples is proved to converge to zero as the number of the samples tends to infinity at each stage of the two algorithms. In addition, the corresponding bounds for the mean-square errors are obtained.
2. MeMBer and CBMeMBer Filters
A Bernoulli RFS has probability of being empty, and probability () of being a singleton whose only element is distributed according to a probability density . The probability density of is
A multi-Bernoulli RFS is a union of a fixed number of independent Bernoulli RFSs , , that is, . is thus completely described by the multi-Bernoulli parameter set with the mean cardinality ββand the probability density [2]:
Throughout this paper, we abbreviate a probability density of the form (2.2) by .
By approximating the multitarget RFS as a multi-Bernoulli RFS at each time step, Mahler proposed the MeMBer recursion, which propagated the multi-Bernoulli parameters of the posterior multitarget density forward in time [2]. The MeMBer filter is summarized as follows.
MeMBer Prediction If at time , the posterior multitarget density is a multi-Bernoulli of the form , then the predicted multitarget density is also a multi-Bernoulli and is given by
where are the parameters of the multi-Bernoulli RFS of births at time :
MeMBer Update If at time , the predicted multitarget density is a multi-Bernoulli of the form ; then the posterior multitarget density can be approximated by a multi-Bernoulli as follows:
where,
By correcting the cardinality bias in the of the MeMBer update step, Vo et al. proposed the CBMeMBer filter [12]. The CBMeMBer recursions are the same as the MeMBer recursions except the update of , which is revised as
Note that not (38) in [12] but (2.10) in our paper is used in the CBMeMBer update step here. The reasons are (1) the (38) in [12] and the (2.10) in our paper are both the approximations of (36) in [12] under the same assumption , but the latter is more precise than former; (2) the (38) in [12] is unbounded at while (2.10) in our paper is bounded at as long as .
For the multi-Bernoulli representation , the probability indicates how likely the th hypothesized track is a true track, and the posterior density describes the distribution of the estimated current state of the track. Hence, denotes the multitarget number and the multitarget state estimate can be obtained by choosing the means or modes from the posterior densities of the hypothesized tracks with existence probabilities exceeding a given threshold.
3. SMC-MeMBer and SMC-CBMeMBer Filters
The SMC implementations of the MeMBer and CBMeMBer recursions are summarized as follows.
SMC-MeMBer and SMC-CBMeMBer Predictions Suppose that at time the (multi-Bernoulli) posterior multitarget density is given and each , , is comprised of a set of weighted samples : Then, given proposal densities and , the predicted (multi-Bernoulli) multitarget density can be computed as follows:
where () is given by birth model; , () and , () are, respectively, given by
SMC-MeMBer and SMC-CBMeMBer Updates Suppose that at time the predicted (multi-Bernoulli) multitarget density is given and each , , is comprised of a set of weighted samples :
Then, the multi-Bernoulli approximation of the SMC-MeMBer-updated multitarget density and SMC-CBMeMBer-updated multitarget density can be computed as follows:
where,
Resampling To reduce the effect of degeneracy, we resample the particles for the multi-Bernoulli parameter set after the update step.
4. Convergence of the Mean-Square Errors for the SMC-MeMBer and SMC-CBMeMBer Filters
To show the convergence results for the SMC-MeMBer and SMC-CBMeMBer filters, certain conditions on the functions need to be met:(1)the transition kernel satisfies the Feller property [18], that is, for all , ;(2)single-sensor/target likelihood density ;(3) are rational-valued random variables such that there exists , some constant , and so that
for all vectors ;(4)the importance sampling ratios are bounded, that is, there exists constants and such that , , and , ;(5)the resampling strategy is multinomial and hence unbiased [19].
First, the convergence of the mean-square errors for the initialization steps of the two filters can easily be established by Lemmaββ0 in [13]. Assuming that at time , we can sample exactly from the initial distribution (). Then, for all ,
hold for some real numbers and which are independent of the number of the sampled particles at time , .
Also, the convergence of the mean-square errors for the resampling steps of the two filters can easily be established by Assumption 5 and Lemmaββ5 in [19].
The main difficulty and greatest challenge is to prove the mean-square convergence for the prediction steps and update steps of the two filters. They are, respectively, established by Propositions 4.1 and 4.2.
Proposition 4.1. Suppose that, for all ,
hold for some real numbers and which are independent of the number of the resampled particles at time , .
Then, after the prediction steps of the SMC-MeMBer and SMC-CBMeMBer filters at time :
hold for a constant and some real numbers and which are independent of , . and are defined by (A.8) and (A.18), respectively. The proof of Proposition 4.1 can be found in Appendix A.1.
Proposition 4.2. Suppose that, for all ,
hold for some real numbers and which are independent of the number of the predicted particles, . Then, after the update steps of the SMC-MeMBer and SMC-CBMeMBer filters at time :
hold for some real numbers , , , , and , which are independent of . , , , and are defined by (A.29), (A.35), (A.47), (A.55), and (A.61), respectively. , denotes the minimum. In addition, , , and depend on the number of targets and decrease with the increase of the target number. From (A.47) and (A.55), it can also be seen that . It indicates that may need more particles than to achieve the same mean-square error bound. The proof of the Proposition 4.2 can be found in Appendix A.2.
Propositions 4.1 and 4.2 show that the bounds for the mean-square error of the SMC-MeMBer and SMC-CBMeMBer prediction steps and update steps at each stage depend on the number of particles. The mean-square errors tend to zero as the number of particles tends to infinity. The bounds for the mean-square errors of these quantities are inversely proportional to the corresponding particle number.
Moreover, from the proofs of Propositions 4.1 and 4.2, it can be seen that(1)Assumptions 1, 3, and 4 ensure that (4.6) holds;(2)Assumption 4 ensures that (4.7) holds;(3)Assumption 2 ensures that (4.12), (4.13), and (4.14) hold;(4)Assumption 5 ensures the convergence of the mean-square errors for the resampling steps of the two filters.
Assumptions 3, 4, and 5 are concerned with the SMC method. They can be satisfied as long as the appropriate sampling strategies are chosen. Assumptions 1 and 2 are concerned with the likelihood and target transition kernel. They may be too restrictive or unrealistic for some practical applications. However, these convergence results give justification to the SMC implementations of the MeMBer and CBMeMBer filters and show how the order of the mean-square errors are reduced as the number of particles increases.
5. Simulations
Here, we briefly describe the application of the convergence results for the SMC-CBMeMBer filter to the nonlinear MTT example presented in Example 1 of [12]. The experiment settings are the same as those of Example 1 except that the number of the particles used for each hypothesized track at time . For convenience, we assume . Assumptions 1β5 are satisfied in this example. So, the SMC-CBMeMBer filter converges to the ground truth in the mean-square sense.
For the SMC-CBMeMBer filter, the estimates of the multitarget number and states, which are derived from the particle multi-Bernoulli parameter set, are unbiased. Therefore, via comparing the tracking performance of the algorithm in the various particle number , the convergence results for the SMC-CBMeMBer filter can be verified to a great extent.
The standard deviation of the estimated cardinality distribution and the optimal subpattern assignment (OSPA) multitarget miss-distance [20] of order with cut-off m, which jointly captures differences in cardinality and individual elements between two finite sets, are used to evaluate the performance of the method. Table 1 shows the time-averaged standard deviation of the estimated cardinality distribution and the time-averaged OSPA in various via 200 MC simulation experiments.
Table 1 shows that both the standard deviation of the estimated cardinality distribution and OSPA decrease with the increase of the particle number . This phenomenon can be reasonably explained by the convergence results derived in this paper: first, the mean-square error of the particle multi-Bernoulli parameter set decreases as the number of the particles increases; then, the more precise estimates of the cardinality distribution and multitarget states can be derived from the more precise particle multi-Bernoulli parameter set, which eventually leads to the results presented in Table 1.
6. Conclusions and Future Work
This paper presents the mathematical proofs of the convergence for the SMC-MeMBer and SMC-CBMeMBer filters and gives the bounds for the mean-square errors. In the linear-Gaussian condition, Vo et al. presented the analytic solutions to the MeMBer and CBMeMBer recursions: GM-MeMBer and GM-CBMeMBer filters [12]. The future work is focused on studying the convergence results and error bounds for the two filters.
Appendix
A.
In deriving the proofs, we use the Minkowski inequality, which states that, for any two random variables and in ,
Using Minkowskiβs inequality, we obtain that, for all,
holds, , for the multi-Bernoulli density and its particle approximation .
Now turn to (4.6). From (2.5), we have
(adding and subtracting a new term)
(using Minkowskiβs inequality)
By Assumption 3 and Lemmaββ1 in [13], we easily obtain that the first term in (A.11) becomes
(since by Assumption 4 and by Assumption 1)
Adding and subtracting a new term in the second term of (A.11), we have
(using Minkowskiβs inequality)
(by (4.4))
where denotes the infimum.
Finally, substituting (A.12) and (A.17) into (A.11), (4.6) is proved with
Now, turn to (4.7). By Lemmaββ0 in [13] and the boundedness of () in Assumption 4, we get that (4.7) holds for a constant . This completes the proof.
Now, turn to (4.12). From (2.9) and (3.10), we have
(adding and subtracting a new term)
(using Minkowskiβs inequality)
(using , )
From (2.12) and (3.14), the expectation in the summation of (A.39) is
(adding and subtracting a new term)
(using Minkowskiβs inequality)
(by , Assumption 2 and (A.4))
(by (4.8), (4.9), and )
From (2.12), and Assumption 2, the denominator of (A.39) is
where is the number of the predicted targets at time .
Substituting (A.44) and (A.43) into (A.39), we get
where , denotes the minimum.
Now, turn to (4.13). First, from (2.12), and Assumption 2, we have
Then, from (2.13), (3.11), and (A.39), we get
(adding and subtracting a new term in the second expectation in the summation)
It holds that (using Minkowskiβs inequality for the second term in the summation)
(by , (A.4), and Minkowskiβs inequality)
(using and Minkowskiβs inequality again for the second term in the summation)
(by , (4.8), (4.9) (A.44), (A.45), and Assumption 2)
Now turn to (4.14). From (2.10) and (3.12), we get
(adding and subtracting a new term)
(using Minkowskiβs inequality)
(by (2.12) and (3.14))
(by (A.39) and (A.47))
Probability density of a Bernoulli random finite set (RFS)
:
Probability density of multi-Bernoulli RFS
:
Abbreviation of . is the multi-Bernoulli parameter set
:
Particle approximation of . denotes that is comprised of the number of the particles
:
Supremum norm. , ββdenotes the supremum
:
Inner product. If the measure in is continuous, it defines the integral inner product; if the measure in is discrete, it defines the summation inner product.
Acknowledgments
This research work was supported by the Natural Science Foundation of China (61004087, 61104051, 61005026), China Postdoctoral Science Foundation (20100481338), and Fundamental Research Funds for the Central University.
References
I. R. Goodman, R. P. S. Mahler, and H. T. Nguyen, Mathematics of Data Fusion, vol. 37 of Theory and Decision Library B, Kluwer Academic Publishers, Norwood, Mass, USA, 1997.
R. Mahler, Statistical Multisource Multitarget Information Fusion, Artech House, Norwood, Mass, USA, 2007.
R. Mahler, βMulti-target Bayes filtering via first-order multi-target moments,β IEEE Transactions on Aerospace and Electronic Systems, vol. 39, no. 4, pp. 1152β1178, 2003.
R. Mahler, βPHD filters of higher order in target number,β IEEE Transactions on Aerospace and Electronic Systems, vol. 43, no. 3, pp. 1523β1543, 2007.
T. Zajic and R. Mahler, βA particle-systems implementation of the PHD multitarget tracking filter,β Proceedings of Spie—the International Society for Optical Engineering, vol. 5096, pp. 291β299, 2003.
H. Sidenbladh, βMulti-target particle filtering for the probability hypothesis density,β in Proceedings of the 6th International Conference on Information Fusion, pp. 800β806, Cairns, Australia, 2003.
B. N. Vo, S. Singh, and A. Doucet, βSequential Monte Carlo methods for multi-target filtering with random finite sets,β IEEE Transactions on Aerospace and Electronic Systems, vol. 41, no. 4, pp. 1224β1245, 2005.
B.-N. VO and W. K. MA, βThe Gaussian mixture probability hypothesis density filter,β IEEE Transactions on Signal Processing, vol. 54, no. 11, pp. 4091β4104, 2006.
B.-T. Vo, B.-N. Vo, and A. Cantoni, βAnalytic implementations of the cardinalized probability hypothesis density filter,β IEEE Transactions on Signal Processing, vol. 55, no. 7, pp. 3553β3567, 2007.
D. E. Clark and J. Bell, βMulti-target state estimation and track continuity for the particle PHD filter,β IEEE Transactions on Aerospace and Electronic Systems, vol. 43, no. 4, pp. 1441β1453, 2007.
W. F. Liu, C. Z. Han, F. Lian, and H. Y. Zhu, βMultitarget state extraction for the PHD filter using MCMC approach,β IEEE Transactions on Aerospace and Electronic Systems, vol. 46, no. 2, pp. 864β883, 2010.
B.-T. Vo, B.-N. Vo, and A. Cantoni, βThe cardinality balanced multi-target multi-Bernoulli filter and its implementations,β IEEE Transactions on Signal Processing, vol. 57, no. 2, pp. 409β423, 2009.
D. E. Clark and J. Bell, βConvergence results for the particles PHD filter,β IEEE Transactions on Signal Processing, vol. 54, no. 7, pp. 2652β2661, 2006.
D. Clark and B.-N. Vo, βConvergence analysis of the Gaussian mixture PHD filter,β IEEE Transactions on Signal Processing, vol. 55, no. 4, pp. 1204β1212, 2007.
P. Cholamjiak, S. Suantai, and Y. J. Cho, βStrong convergence to solutions of generalized mixed equilibrium problems with applications,β Journal of Applied Mathematics, vol. 2012, Article ID 308791, 18 pages, 2012.
V. Colao, βOn the convergence of iterative processes for generalized strongly asymptotically -pseudocontractive mappings in Banach spaces,β Journal of Applied Mathematics, vol. 2012, Article ID 563438, 18 pages, 2012.
G. Gu, S. Wang, and Y. J. Cho, βStrong convergence algorithms for hierarchical fixed points problems and variational inequalities,β Journal of Applied Mathematics, vol. 2011, Article ID 164978, 17 pages, 2011.
C. Zhu and G. Yin, βFeller continuity, recurrence, and stabilization of regime-switching diffusions,β in Proceedings of the 49th Conference on Decision and Control (CDC '10), pp. 667β672, Atlanta, Ga, USA, 2010.
D. Crisan and A. Doucet, βA survey of convergence results on particle filtering methods for practitioners,β IEEE Transactions on Signal Processing, vol. 50, no. 3, pp. 736β746, 2002.
D. Schuhmacher, B.-T. Vo, and B.-N. Vo, βA consistent metric for performance evaluation of multi-object filters,β IEEE Transactions on Signal Processing, vol. 56, no. 8, pp. 3447β3457, 2008.