Research Article  Open Access
A New Sparse GaussHermite Cubature Rule Based on RelativeWeightRatios for BearingRanging Target Tracking
Abstract
A new sparse GaussHermite cubature rule is designed to avoid dimension explosion caused by the traditional full tensorproduct based GaussHermite cubature rule. Although Smolyak’s quadrature rule can successfully generate sparse cubature points for high dimensional integral, it has a potential drawback that some cubature points generated by Smolyak’s rule have negative weights, which may result in instability for the computation. A relativeweightratio criterion based sparse GaussHermite rule is presented in this paper, in which cubature points are kept symmetric in the input space and corresponding weights are guaranteed to be positive. The generation of the new sparse cubature points set is simple and meaningful for practice. The difference between our new sparse GaussHermite cubature rule and other cubature rules is analysed. Simulation results show that, compared with Kalman filter with those types of full tensorproduct based GaussHermite rules, our new sparse GaussHermite cubature rule based Kalman filter can lead to a substantially reduced number of cubature points, more stable computation capability, and maintaining the accuracy of integration at the same time.
1. Introduction
Bayesian recursive estimation is commonly used in target tracking, positioning, and signal processing [1]. Generally, a Bayesian recursive estimation algorithm requires a state model and a measurement model. The a posteriori density function can describe the behaviour of the estimated state [2]. Since a closed form solution to the Bayesian recursive estimation is available only for a few special cases [3], such as the linear Gaussian system (which leads to the classical standard Kalman filter), a suboptimal solution is a preferable choice in the general case [4, 5]. Ito and Xiong [4] suggest that a local Gaussian filter can be used to approximate the general Bayesian recursive estimation suboptimally. The core problem of local Gaussian filters is in fact a high dimensional Gaussian weighted integration, which has been studied in numerical analysis; see [6–9] and the references therein. Lots of researches concentrate on the numerical approximation methods to solve the Gaussian weighted integral problem, and different approaches result in different local Gaussian filters, such as the cubature Kalman filter [10], the quadrature Kalman filter [11], and related variants [12–14]. GaussHermite filter, introduced by [2, 4], makes use of the GaussHermite quadrature rule and has the highest accuracy among all the above filters. However, it suffers from the curse of dimensionality since the number of cubature points increases exponentially with the dimension of the state.
To avoid the curse of dimensionality, several sparse rules are developed. Smolyak rule [15] is one of the useful tools to generate a small number of cubature points for high dimensional integral. The computational cost does not increase exponentially by using this sparse grid method. Reference [16] compares the approximation accuracies of various sparse grid methods, including trapezoidal rule, Crenshaw Curtis rule, and Gauss Patterson rule. Reference [17] combines Smolyak’s rule and Gaussian Quadrature rule for high dimensional likelihood function in economic models. Jia et al. [18, 19] make a tracking comparison among Smolyak’s rule based GaussHermite filter of different levels and the traditional cubature filters in the context of determination of the spacecraft attitude and the lowerearth orbit satellite orbit. Simulations prove the effectiveness of Smolyak’s rule based GaussHermite filter. Reference [20] proposes the multiple sparse grid GaussHermite filter based on sparse grid GaussHermite filter and statespace partitioning. It is claimed that the computational burden is further reduced with respect to the GaussHermite filter and the sparse grid GaussHermite filter.
However, as Heiss and Winschel [17] mentioned, there exists a potential drawback that some of the cubature points have negative weights, which may result in instability for computation. So we come up with the idea to avoid this potential drawback. In this paper, a classification of full tensorproduct GaussHermite cubature points is obtained by using equivalence classification by position permutation and signum function. There exist plenty of cubature points with low weights, which can be deleted from the cubature scheme directly considering a relativeweightratio threshold. Comparing with the Smolyakbased rule, our construction of the new GaussHermite rule is simple and practically meaningful. The corresponding weights of the cubature points are all positive; meanwhile, the full symmetry property still remains. There also exist some interesting relationships among the GaussHermite filter, the 3rd embedding cubature filter, the sparse GaussHermite filter, and the Unscented Kalman filter.
The remainder of this paper is organized as follows: In Section 2, a brief review of the nonlinear system and its Bayesian recursive estimation framework is presented. In Section 3, a cycle of general local Gaussian filter is presented, which offers six kinds of Gaussian weighted integrals. This leads to various specific local Gaussian filters by various choice of the cubature points set and the corresponding weights. In Section 4, a full tensorproduct GaussHermite integral cubature rule is presented. By introducing a simple and elegant operator, the full tensorproduct based cubature points can be divided into different categories. A small set of positive weighted cubature points is generated by a threshold of relativeweightratio. Relationships between the new sparse GaussHermite filter and the traditional cubature filters are analysed as well. In Section 5, a typical BearingRanging tracking problem with a general 2D manoeuvring target motion is demonstrated to test the performance of our new sparse GaussHermite filter. Some conclusions are given in Section 6.
2. Bayesian Recursive Estimation Algorithm
Consider the following discrete nonlinear system:
where (1) and (2) are the motion model and the measurement model of the system, respectively; is the state of the system and is the measurement; is the process noise and is the measurement noise. is independent with and . By the discrete time ChapmanKolmogoroff equation,where . By the Bayesian formula,where . Equations (3) and (4) are the time update formula and the measurement update formula. Figure 1 shows the recursive Bayesian estimation algorithm. When is obtained, we can estimate the state and its covariance matrix by minimizing the mean square error. The results are presented as follows:However, when becomes an arbitrary probability density distribution, it is difficult to calculate the high dimensional integrations in (5) directly. A suboptimal way is to place normal distributions on , , and ; thus the firstorder moment and secondorder moment can be used to describe the posterior probability. In other words, at time epoch , the posterior density isAt time epoch , the predicted probability density of the state isCorrespondingly, the posterior density of the state is
3. Local Gaussian Filter
Section 2 points out that, under the Gaussian assumption, the firstorder moment and secondorder moment can be used to describe the property of the state. Hence, a suboptimal local Gaussian filter algorithm can replace the Bayesian recursive estimation (3) and (4). Details of general Gaussian filter can be found in [4].(1)Time update.(i)Prediction of state :(ii)Prediction of covariance matrix given state :(2)Measurement update.(i)Prediction of measurement :(ii)Covariance matrix given : (iii)Covariance matrix between state and measurement :(iv)Gain of the filter:(v)State update and its covariance matrix :
We can see that (9), (10), (14), and (15) are similar to those corresponding formulae of standard Kalman filter. The only difference between local Gaussian filter and standard Kalman filter is the calculation of , , , , and . For standard Kalman filter, there exist specific formulae, while for the local Gaussian filter, numerical approximation methods need to calculate the higher dimensional integrals. A general high dimensional integral with Gaussian type weighting function can be expressed as follows:Suppose that , where is a lower triangular matrix. By a linear transform , we can get thatVarious methods can approximate (17), such as the spherical radial integral rule [7, 21], GaussHermite rule [2], central differential rule [4], Stochastic Integral Rule [22], Gaussian localized cumulative distribution rule [23], and Smolyak rule based sparse grid rule [18]. All of them use specific ways to generate and their corresponding weights , where is the number of cubature points set.
Remark 1. Generally, the choice of has no relationship with the estimation of state and its covariance matrix , but uniquely determined by the specific cubature integral rule. In fact, almost all cubature rules based can be computed and stored offline.
4. Sparse GaussHermite Integral Rules
4.1. Symmetry Set Based on Full Position Permutation
Definition 2. In order to utilize the symmetry property of the cubature points, is introduced to achieve a more concise denotation of the cubature points set. Denote as a nonnegative seed; we have its corresponding seed vector as a dimensional vector , where is the state dimension. Then we have , where means that the set of elements consists of the full permutation vectors of the seed vector and is generated by randomly setting one or several nonzero values to their opposites in .
Let the state dimension be and take as an example, where is the seed. Since two values are given, then the dimensional seed vector is . It has total full permutation, so has elements. For simplicity, suppose ; then has elements, which is presented as the following set:
Remark 3. With the help of full position permutation, it becomes much easier to describe the cubature points set . Furthermore, it brings some interesting properties for the cubature points.
4.2. Full TensorProduct Based GaussHermte Integral Rule
The classical onedimension GaussHermite rule can be easily obtained according to [24]. Table 1 presents us with cubature points set and corresponding weights from orders to . Here the order also represents the number of cubature points in each dimension. We can see that the points and weights are both symmetric in . For details about the construction of GaussHermite quadrature points, one can refer to [25]. Furthermore, [26] provides a better and more stable way to construct cubature points and weights than moment matching method.
For dimension integral problem, a full tensorproduct based GaussHermite rule can be obtained directly by the following formula [2]:As mentioned above, is the dimension of the state, is the number of the cubature points for each dimension, and as presented in (19). Each point of the cubature set is a dimensional vector; that is, . The weight of point can be calculated as . For example, let and ; we can easily get the cubature points and the corresponding weights as follows:Recall Definition 2; the cubature points in (20) can be categorized as three sets ; that is,and the weights can be summarized as three types as well: The number of each category can be easily computed asThe total number of cubature points . Equations (21), (22), and (23) show that such a classification is an equivalent notation of the full tensorproduct rule. Now suppose that remains unchanged; we would like to consider the general case of dimension. According to Definition 2, there exist categories. The corresponding , , and the number of each category are shown as follows:
The total number of cubature points . It also shows that such an expression is another form of the full tensorproduct based rule. When the dimension of state increases, the number of cubature points increases exponentially with . The curse of dimensionality limits the usage of GaussHermite integral rule.
Now, we are going to give the cubature points set and their corresponding weights with . It can be easily conducted that when , there is only one cubature point; that is, , and its weight is . When , we can refer to Table 1 to find that there is only one nonnegative value (i.e., 1), so there exists only one category as well: When , it is interesting to find that the GHKF is similar to the 3rd embedding cubature Kalman filter [27]. We will elaborate their connections in the following Remark.
Remark 4. The main difference between GHKF and the 3rd embedding cubature Kalman filter is the central point and its weight utilized. The classical standard 3rd cubature rule based on spherical radial rule [10, 28], whose cubature points set is , may exceed the range of state. Both the GHKF and the 3rd embedding cubature rule are stable for high dimensional integral rule. However, since the number of cubature points in 3rd embedding cubature Kalman filter is , we still face the threat of dimension explosion when dealing with high dimensional system.
When , there still exists type of category which is analogous to the case . Here we can simply regard as of case only to remember that is positive. We have and according to Table 1. The total number of cubature points becomes . More specifically, the cubature points set , the corresponding weights , and the number for each category are shown as follows:where . It can be seen that the number of each cubature category has a factor , which will result in the dimension explosion as well.
When , the total number of cubature points is . It can be classified as follows:The corresponding weights are presented aswhere . , . We can calculate the number of each category to obtain the following equations:Here we can see that, for , the number of each category has a factor , so those cubature points sets are not a good choice for the high dimensional system. However, the number of each category is a polynomial of the dimension of state when . In the face of the dimension explosion, it is rather tempting that we can just choose several categories to approximate the high dimensional integration. We have introduced the relativeweightratio as a criterion to achieve this goal. Detailed analysis is presented in the following subsection.
4.3. Sparse GaussHermite Rule Based on RelativeWeightRatio
Take as an example and recall (21) and (22) which present the cubature points set and corresponding weights set, respectively. The relativeweightratio between category and category can be easily obtained:Equation (30) shows that as the subindex of the cubature points set increases, the decreases exponentially. It indicates that when considering a certain degree of accuracy, there is no need to use all GaussHermite cubature points. When a threshold value is settled, only the categories satisfying will be sufficient for the high dimensional integral. For example, set ; then we haveAccording to the criterion, the categories including , , , and are enough to approximate the integral. The total number of the cubature points isThe number of cubature points has a polynomial relationship with the dimension of state. If we set threshold value as , the remaining categories will become , , and . The corresponding number of the cubature points is
Remark 5. By setting the ratio threshold , most low weighted cubature points will be eliminated. Such a strategy can efficiently eliminate the problem of dimension explosion when dealing with high dimensional system. Bear in mind that since lots of low weighted cubature points are eliminated, the sum of remaining points’ weights does not equal to 1. We can deal with this problem by renormalizing the remaining weights and meanwhile maintaining the relativeweightratio. Unlike the Unscented Transform rules [29] or the higher order spherical radial rule cubature rule [28], our transformation keeps all weights positive.
Remark 6. When , considering and as selected cubature points, the total number is which equals the number of the UKF with parameter [29]. The difference lies in the determination of weights of the cubature points. In the context of sparse GHKF, the weight of is . However under the background of UKF, the weight becomes . Similarly, we have weights of as and for sparse GHKF and UKF, respectively. When dimension , the weight of becomes negative in the UKF context, which is a potential unstable factor for integral approximation. The sparse GaussHermite cubature points for can also be obtained by repeating the above discussed process with a preset relativeweightratio. Comparing with , the weight of the former category is larger than the latter one.
5. Numerical Investigation
In this section, a Nearly Constant Turning (NTC) manoeuvring target model is settled to verify the tracking performance of our new sparse GaussHermite Kalman filter. NCT model is commonly used in aircraft flying control. We consider the 2dimensional movement. Suppose the unknown turning rate is . Since is unknown, the NCT model is a timevarying nonlinear model [10]. The discrete state model can be settled as follows:where . and represent the position and velocity of axis , respectively. and represent the position and velocity of axis , respectively. is the sampling time interval. is the turning rate at time epoch . The corresponding process noise , where can be rewritten as follows:
and represent the Power Spectrum Density for positions and . is the Power Spectrum Density of the turning rate. Suppose that the sensor platform is settled at the origin points. BearingRanging measurement is used for the target tracking. The measurement equation can be expressed as follows:where is the measurement noise, which satisfies the Gaussian distribution with zero mean. The corresponding covariance matrix can be expressed as follows:The corresponding parameters are settled by Table 2 [10].

We set the initial state of the target as Then the corresponding covariance matrix of the initial state isThe estimate of initial state of the filter is randomly generated by the distribution . Our new sparse GaussHermite Kalman filter with the cubature sets is used in the experiment. In order to make a reliable comparison between different filters, we implement our experiment with the (5th) CKF [28], full tensorproduct based GaussHermite Kalman filter () [2], and Smolyak rule based sparse GaussHermite Kalman filter [18] .
The above experiment is repeated 100 times independently. Root mean square errors , , and are used to evaluate the tracking performance of the position, velocity, and the turning rate. The detailed formulations are shown as follows: Similarly, CRMSE are used to evaluate the average tracking performance:
where represents the th estimate element of state at time epoch for the th experiment. represents the th theoretical element of state at time epoch for the th experiment. represents the Monte Carlo experiment times. represents the state number of each experiment. We set . in this experiment. For a single simulation experiment, the tracking performance is shown in Figures 2, 3, and 4. For repeated experiments, the tracking performance is shown in Figures 5, 6, and 7. The CRMSEs of different filters are shown in Table 3.

From Table 3, it can be seen that the full tensorproduct based GaussHermite filter and the (5th) CKF have almost exactly the same tracking performance, while the Smolyakbased sparse GaussHermite filter has relatively lower tracking accuracy. The new sparse GaussHermite filter has better tracking accuracy comparing with sparse GaussHermite filter and has less tracking accuracy comparing with the full tensorproduct based GaussHermite filter and the (5th) CKF. Our new sparse GaussHermite filter is a good choice to track this manoeuvring target problem with a relative sparse cubature points set where all corresponding weights are kept positive. Furthermore, our method burdens low computational costs and maintains stable numerical accuracy.
6. Conclusion and Discussion
GaussHermite cubature rule is an effective way to approximate nonlinear Gaussian weighted type integral. However, the full tensorproduct based GaussHermite cubature rule may cause a dimension explosion and is rarely helpful for high dimensional case. Smolyak rule based sparse GaussHermite cubature rule is a way to generate small cubature set by specific generation design, which has a potential drawback that some cubature points may have negative weights. This novel GaussHermite rule is easily generated by using our notation in Definition 2 to classify the full tensorproduct based GaussHermite cubature points. Based on a proper relativeratioweight criterion, a sparse cubature set is regenerated. Simulations show that our sparse GaussHermite rules will have comparable accuracy with the full tensorproduct based GaussHermite filter and higher precision comparing with Smolyak’s rule based GaussHermite filter. Furthermore, our method can lead to a substantially reduced number of cubature points and a more stable high dimensional integration. Our method can maintain accuracy of integration at the same time and could be a good choice for practice problems in engineering.
Disclosure
The views and opinions expressed in this paper are those of the authors and do not necessarily reflect the official policy or position of the iGMAS.
Conflicts of Interest
The authors declare that there are no conflicts of interest regarding the publication of this paper.
Acknowledgments
This work is supported by the iGMAS (international GNSS Monitoring and Assessment Service) and the Natural Science Foundation of China (no. 61573367).
References
 F. Gustafsson, F. Gunnarsson, N. Bergman et al., “Particle filters for positioning, navigation, and tracking,” IEEE Transactions on Signal Processing, vol. 50, no. 2, pp. 425–437, 2002. View at: Publisher Site  Google Scholar
 I. Arasaratnam, S. Haykin, and R. J. Elliott, “Discretetime nonlinear filtering algorithms using GaussHermite quadrature,” Proceedings of the IEEE, vol. 95, no. 5, pp. 953–977, 2007. View at: Publisher Site  Google Scholar
 M. Šimandl and J. Duník, “Derivativefree estimation methods: new results and performance analysis,” Automatica. A Journal of IFAC, the International Federation of Automatic Control, vol. 45, no. 7, pp. 1749–1757, 2009. View at: Publisher Site  Google Scholar  MathSciNet
 K. Ito and K. Xiong, “Gaussian filters for nonlinear filtering problems,” Institute of Electrical and Electronics Engineers. Transactions on Automatic Control, vol. 45, no. 5, pp. 910–927, 2000. View at: Publisher Site  Google Scholar  MathSciNet
 Y. Wu, D. Hu, M. Wu, and X. Hu, “A numericalintegration perspective on gaussian filters,” IEEE Transactions on Signal Processing, vol. 54, no. 8, pp. 2910–2921, 2006. View at: Publisher Site  Google Scholar
 A. H. Stroud and D. D. Stancu, “Quadrature formulas with multiple Gaussian nodes,” Journal of the Society for Industrial and Applied Mathematics Series B Numerical Analysis, vol. 2, no. 1, pp. 129–143, 1965. View at: Google Scholar  MathSciNet
 A. Genz, “Fully symmetric interpolatory rules for multiple integrals,” SIAM Journal on Numerical Analysis, vol. 23, no. 6, pp. 1273–1283, 1986. View at: Publisher Site  Google Scholar  MathSciNet
 A. Genz and B. D. Keister, “Fully symmetric interpolatory rules for multiple integrals over infinite regions with Gaussian weight,” Journal of Computational and Applied Mathematics, vol. 71, no. 2, pp. 299–309, 1996. View at: Publisher Site  Google Scholar  MathSciNet
 A. Genz, “Fully symmetric interpolatory rules for multiple integrals over hyperspherical surfaces,” Journal of Computational and Applied Mathematics, vol. 157, no. 1, pp. 187–195, 2003. View at: Publisher Site  Google Scholar  MathSciNet
 I. Arasaratnam and S. Haykin, “Cubature Kalman filters,” Institute of Electrical and Electronics Engineers. Transactions on Automatic Control, vol. 54, no. 6, pp. 1254–1269, 2009. View at: Publisher Site  Google Scholar  MathSciNet
 S. Bhaumik and Swati, “Cubature quadrature Kalman filter,” IET Signal Processing, vol. 7, no. 7, pp. 533–541, 2013. View at: Publisher Site  Google Scholar  MathSciNet
 A. K. Singh and S. Bhaumik, “Cubature quadrature Kalman filter for maneuvering target tracking,” in Proceedings of the 1st IEEE International Conference on Smart Sensors and Application, ICSSA 2015, pp. 24–29, May 2015. View at: Publisher Site  Google Scholar
 A. K. Singh and S. Bhaumik, “Higher degree cubature quadrature kalman filter,” International Journal of Control, Automation and Systems, vol. 13, no. 5, pp. 1097–1105, 2015. View at: Publisher Site  Google Scholar
 A. K. Singh, S. Bhaumik, and P. Date, “Quadrature filters for onestep randomly delayed measurements,” Applied Mathematical Modelling. Simulation and Computation for Engineering and Environmental Systems, vol. 40, no. 1920, pp. 8296–8308, 2016. View at: Publisher Site  Google Scholar  MathSciNet
 S. Smolyak, “Quadrature and interpolation formulas for tensor products of certain classes of functions,” Soviet Mathematics, Doklady, vol. 4, no. 5, pp. 240–243, 1963. View at: Google Scholar  MathSciNet
 T. Gerstner and M. Griebel, “Dimensionadaptive tensorproduct quadrature,” Computing. Archives for Scientific Computing, vol. 71, no. 1, pp. 65–87, 2003. View at: Publisher Site  Google Scholar  MathSciNet
 F. Heiss and V. Winschel, “Likelihood approximation by numerical integration on sparse grids,” Journal of Econometrics, vol. 144, no. 1, pp. 62–80, 2008. View at: Publisher Site  Google Scholar  MathSciNet
 B. Jia, M. Xin, and Y. Cheng, “Sparse gausshermite quadrature filter with application to spacecraft attitude estimation,” Journal of Guidance, Control, and Dynamics, vol. 34, no. 2, pp. 367–379, 2011. View at: Publisher Site  Google Scholar
 B. Jia, M. Xin, and Y. Cheng, “Sparsegrid quadrature nonlinear filtering,” Automatica. A Journal of IFAC, the International Federation of Automatic Control, vol. 48, no. 2, pp. 327–341, 2012. View at: Publisher Site  Google Scholar  MathSciNet
 R. Radhakrishnan, A. K. Singh, S. Bhaumik, and N. K. Tomar, “Multiple sparsegrid GaussHermite filtering,” Applied Mathematical Modelling. Simulation and Computation for Engineering and Environmental Systems, vol. 40, no. 78, pp. 4441–4450, 2016. View at: Publisher Site  Google Scholar  MathSciNet
 A. H. Stroud, “Some cubature formulas for the surface of the $n$sphere,” SIAM Journal on Numerical Analysis, vol. 10, no. 4, pp. 559–569, 1973. View at: Publisher Site  Google Scholar  MathSciNet
 A. Genz and J. Monahan, “Stochastic integration rules for infinite regions,” SIAM Journal on Scientific Computing, vol. 19, no. 2, pp. 426–439, 1998. View at: Publisher Site  Google Scholar  MathSciNet
 J. Steinbring and U. D. Hanebeck, “LRKF revisited: the smart sampling Kalman filter (S^{2}KF),” Journal of Advances in Information Fusion, vol. 9, no. 2, pp. 106–123, 2014. View at: Google Scholar
 W. Press, B. Flannery, S. Teukolsky, and W. Vetterling, Numerical Recipes in C: The Art of Scientific Computing, Cambridge University Press, 1999.
 G. H. Golub and C. F. Van Loan, Matrix Computations, The Johns Hopkins University Press, Baltimore, Md, USA, 3rd edition, 1996. View at: MathSciNet
 G. H. Golub and J. H. Welsch, “Calculation of Gauss quadrature rules,” Mathematics of Computation, vol. 23, no. 106, article 221, 1969. View at: Publisher Site  Google Scholar  MathSciNet
 X.C. Zhang, “A novel cubature kalman filter for nonlinear state estimation,” in Proceedings of the 52nd IEEE Conference on Decision and Control, CDC 2013, pp. 7797–7802, December 2013. View at: Publisher Site  Google Scholar
 B. Jia, M. Xin, and Y. Cheng, “Highdegree cubature Kalman filter,” Automatica. A Journal of IFAC, the International Federation of Automatic Control, vol. 49, no. 2, pp. 510–518, 2013. View at: Publisher Site  Google Scholar  MathSciNet
 S. J. Julier and J. K. Uhlmann, “Unscented filtering and nonlinear estimation,” Proceedings of the IEEE, vol. 92, no. 3, pp. 401–422, 2004. View at: Publisher Site  Google Scholar
Copyright
Copyright © 2017 Lijun Peng et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.