sensors

Communication
Robust Cubature Kalman Filter for Moving-Target Tracking with
Missing Measurements
Samer Sahl 1,2 , Enbin Song 1, * and Dunbiao Niu 3

1 College of Mathematics, Sichuan University, Chengdu 610065, China; samersa7l@aun.edu.eg

2 Department of Statistics, Assiut University, Assiut 71515, Egypt
3 College of Electronic and Information Engineering, Tongji University, Shanghai 201804, China;
dunbiaoniu_sc@163.com
* Correspondence: e.b.song@163.com

Abstract: Handling the challenge of missing measurements in nonlinear systems is a difficult problem
in various scientific and engineering fields. Missing measurements, which can arise from technical
faults during observation, diffusion channel shrinking, or the loss of specific metrics, can bring
many challenges when estimating the state of nonlinear systems. To tackle this issue, this paper
proposes a technique that utilizes a robust cubature Kalman filter (RCKF) by integrating Huber’s
M-estimation theory with the standard conventional cubature Kalman filter (CKF). Although a
CKF is often used for solving nonlinear filtering problems, its effectiveness might be limited due
to a lack of knowledge regarding the nonlinear model of the state and noise-related statistical
information. In contrast, the RCKF demonstrates an ability to mitigate performance degradation
and discretization issues related to track curves by leveraging covariance matrix predictions for state
estimation and output control amidst dynamic disruption errors—even when noise statistics deviate
from prior assumptions. The performance of extended Kalman filters (EKFs), unscented Kalman filters
(UKFs), CKFs, and RCKFs was compared and evaluated using two numerical examples involving the
Univariate Non-stationary Growth Model (UNGM) and bearing-only tracking (BOT). The numerical
experiments demonstrated that the RCKF outperformed the EKF, EnKF, and CKF in effectively
handling anomaly errors. Specifically, in the UNGM example, the RCKF achieved a significantly
lower ARMSE (4.83) and ANCI (3.27)—similar outcomes were observed in the BOT example.

Keywords: cubature Kalman filter; missing data; robust cubature Kalman filter
Citation: Sahl, S.; Song, E.; Niu, D.
Robust Cubature Kalman Filter for
Moving-Target Tracking with Missing
Measurements. Sensors 2024, 24, 392. 1. Introduction
https://doi.org/10.3390/s24020392 Since its development in 1960 by Kalman [1], the Kalman filter has been widely utilized
Academic Editor: Paolo Mercorelli in ocean-atmosphere science to develop numerous nonlinear filters [2]. The EKF, UKF,
EnKF, and CKF are commonly used variations of the Kalman filter [3]. The EKF linearizes
Received: 30 October 2023 nonlinear systems using the Jacobian matrix and first-order Taylor expansion, making it
Revised: 2 December 2023
suitable for navigation, target tracking, data fusion, and state estimation [4]. However,
Accepted: 15 December 2023
the Jacobian matrix has limitations in achieving precise linearization with decreasing
Published: 9 January 2024
gradients [5]. The UKF, on the other hand, utilizes the unscented transform to avoid the
need for computing the Jacobian matrix. However, it requires accurate prior knowledge
of the system noise statistics, which can be challenging to describe correctly in dynamic
Copyright: © 2024 by the authors.
environments, potentially leading to incomplete or divergent filtering solutions [6].
Licensee MDPI, Basel, Switzerland. The EnKF belongs to the class of particle filters, where an ensemble of state estimates
This article is an open access article is selected to represent the initial probability distribution [7,8]. These estimates are propa-
distributed under the terms and gated through the nonlinear system, approximating the probability density function of the
conditions of the Creative Commons true state [9]. However, for highly nonlinear applications requiring high precision and a
Attribution (CC BY) license (https:// finite ensemble size, the EnKF may not be optimal [10]. The CKF utilizes a third-degree
creativecommons.org/licenses/by/ cubature rule and offers advantages such as reduced parameters [11,12], improved sta-
4.0/). bility, and accuracy compared to the UKF [13,14]. It is widely used to handle nonlinear

Sensors 2024, 24, 392. https://doi.org/10.3390/s24020392 https://www.mdpi.com/journal/sensors

Sensors 2024, 24, 392 2 of 18

problems [15], but applying the CKF to a nonlinear system requires knowledge of the
mathematical model and noise statistics, which can be challenging to obtain in practical
applications [16].
Over the last few years, considerable effort has been spent on developing the RCKF
based on Huber’s idea of M-estimation and the traditional CKF. It can handle the problem
of performance degradation, and the tracking curves are discretized whenever the data
diverge from the previous noise statistics. Table 1 shows a comparison of relevant works
for illustration.

Table 1. Comparison of relevant works.

Authors Method Results

Huber’s M-estimation-based Simulations demonstrated the
robust CKF and robust square superior performance in terms
root CKF adapted to of accuracy, robustness,
Tiancheng Li et al. [17]
anomalous measurement and reliability compared to
noise using innovation standard methods for
covariance comparison. target tracking.
Robust adaptive CKF to The experiment confirmed the
reduce kinematic model errors proposed strategy’s
Zhao et al. [16] through covariance effectiveness in dynamic
adjustment and dynamic systems with high dynamics
disturbance processing. and weak signals.
RCKF enhanced GNSS/INS Numerical experiments and
accuracy in GNSS-denied field tests demonstrated the
Cui Bingbo et al. [18] environments by considering RCKF’s superior robustness
noise using missing compared to the CKF
observations. and EKF.
Adaptive robust CKF
(ARCKF) based on the Simulations favored the
H-infinity CKF by recommended approach over
Xiangzhou Ye et al. [19] incorporating two adaptable the HCKF for handling model
algorithm components to errors and
address erroneous system aberrant observations.
models and noise statistics.

Previous studies have shown that the RCKF algorithm provides significant improvements
in tracking accuracy and stability in many applications, outperforming traditional methods.
However, these studies failed to deal with the problem of missing measurements in nonlinear
systems, which frequently occur in practical scenarios due to imprecise observations.
This paper suggests an RCKF technique based on Huber’s idea of M-estimation and
the CKF for the estimation of the state of nonlinear systems with missing measurements.
Missing measurements are often an inescapable occurrence in many practical scenarios
due to the specific variables associated with erroneous observations. Interruptions in
the technical aspects of observation, shrinkage occurrences in the diffusion channels,
and erroneously lost measurements are some of the reasons for missing data. In addition,
data inaccessibility is also a possibility [20,21]. To describe missing measurements using
random variables, the Bernoulli distribution is more commonly used than the Markov
chain [22]. We summarize the contributions and significance of this paper as follows:
1. The RCKF was developed by integrating Huber’s M-estimation theory with the
standard CKF to effectively handle nonlinear systems, with missing measurements
characterized using random variables following the Bernoulli distribution.
2. The RCKF exhibited superior performance compared to the EKF, EnKF, and CKF
in terms of accuracy and reliability on two moving-target tracking models (UNGM
and BOT) with missing measurements, indicating that the RCKF is the most effective
approach for nonlinear systems with missing measurements.
Sensors 2024, 24, 392 3 of 18

This paper is organized as follows: Section 2 presents an analysis of a nonlinear system

that is afflicted by missing measurements. Section 3 delves into the CKF and provides an
in-depth discussion on Gaussian Bayesian filters. Section 4 proposes a novel approach to
enhance the robustness of the CKF using robust estimation theory. In Section 5, a robust
CKF is presented and applied in two numerical instances to compare its performance with
existing filters. Finally, Section 6 concludes the paper.

2. Nonlinear System with Missing Measurements

The following equations formulate a nonlinear system with missing measurements [22]:

λk = g(λk−1 ) + δk−1 , (1)

θk = ψk h(λk ) + ξ k , (2)

where k is the discrete time index; λk ∈ Rn is the state vector; θk ∈ Rm is the measurement
vector; δk−1 ∈ Rn and ξ k ∈ Rm are process noise and measurement noise, respectively; g(·)
and h(·) are the known nonlinear functions. Additionally, the nonlinear systems (1) and (2)
are assumed to have the following properties:
1. The initial state follows a Gaussian distribution, i.e., λ0 ∼ N (λ̄0 , B0 ).
2. The noise sequences δk−1 and ξ k are independent Gaussian sequences with zero
means, and the covariance matrix of δk−1 is denoted as qk−1 , while the covariance
matrix of ξ k is denoted as rk .
A Bernoulli distribution is utilized to describe missing measurements by incorporating
the measurement function ψk with the following property-related statistical features: p(ψk = 1)
= E(ψk ) = p and p(ψk = 0) = E(ψk ) = 1 − p [23]. When ψk = 1, the sensor obtains data
with precision; conversely, it simply captures noise when ψk = 0, and no measurements are
taken. Note that when referring to models (1) and (2) as reflective of the existence of missing
measurements, the system receives data from the sensor at all times, and it is impossible
to determine whether the data θk are obtained when ψk = 1 or ψk = 0. Despite the fact
that the nonlinear system with missing measurements has become increasingly prevalent
in real-life situations owing to multiplicative noise ψk , it complicates the attainment of
optimal filtering outcomes.
This study aimed to utilize the concept of the least mean square error to construct
an RCKF for nonlinear discrete systems represented by (1) and (2). The RCKF method is
dependent on the robust M-estimation technique.
Bayesian filtering seeks to estimate the probability density function (PDF) of state
variable λk based on the sequence of all available measurements Θk−1 = {θ1 , θ2 , · · · , θk−1 }
up to time k. Thus, it is required to construct the posterior PDF p(λk |Θk ) and the prior PDF
p(λk |Θk−1 ) of the state variables λk . That is, the condition PDF of λk given Θk and Θk−1
can be recursively computed using the provided solutions.
Z
p ( λ k | Θ k −1 ) = p(λk |λk−1 ) p(λk−1 |Θk−1 )dλk−1 , (3)

p ( θ k | λ k ) p ( λ k | Θ k −1 )
p(λk |Θk ) = , (4)
p ( θ k | Θ k −1 )

Assuming that p(λk−1|k−1 |Θ) ≈ N(λ̂k−1|k−1, Bk−1|k−1 ) and p(λk|k−1 |Θ) ≈ N(λ̂k|k−1, Bk|k−1 ),
we can obtain the conditional probability densities in (3) and (4) by calculating the mean and
covariance using the Kalman filter (KF) [24]. The KF has two stages of operation: time and
measurement updates. While some sources use the terms “forecast” and “analysis”, others
use “prediction” and “update” to describe these two stages. For details, see [25]; Figure 1
summarizes the algorithm of the KF.
Sensors 2024, 24, 392 4 of 18

Figure 1. The algorithm of the KF.

3. Cubature Kalman Filter with Missing Measurements

Since the calculation of multivariate integrals is difficult, utilizing approximate meth-
ods is essential [16]. The CKF, invented by Arasaratnam et al. in 2009, is a Bayesian
filter that provides an approximation for nonlinear filtering problems at the discrete time
scale. It assumes that the predictive density of the combined state measurement follows a
Gaussian distribution. The CKF employs the third-degree spherical-radial cubature rule to
numerically compute integrals, scaling points linearly with the state vector dimension [15].
It effectively addresses complex nonlinear problems with high dimensions [26].
The classical form of the CKF is introduced below, comprising two distinct components—
the measurement update and the time prediction [27]:
1. The time prediction is as follows:
I The posterior probability distribution of a given k − 1 time is

p(λk−1 |Θk−1 ) = N (λk−1 : λ̂k−1|k−1 , B̂k−1|k−1 ). (5)

By Cholesky decomposition, The expression denoting the error covariance at

time k − 1, denoted as Bk−1|k−1 , is given by

Bk−1|k−1 = Ak−1|k−1 AkT−1|k−1 , (6)

where Ak−1|k−1 denotes a diagonal time k − 1 matrix.

II Calculating the cubature points.

lambdat,k−1|k−1 = Ak−1|k−1 ζ t + λ̂k−1|k−1 , (7)

where λt,k−1|k−1 (t = 1, 2, · · · , 2n) represents the system state of the t-th cubature
point at time k − 1. The cubature points set is denoted as [ζ t ] and can be defined as
r
2n
ζt = ⟨1⟩ t t = 1, 2, · · · , 2n, (8)
2
Sensors 2024, 24, 392 5 of 18

where 2n is the number of cubature points or twice the state dimension;

⟨1⟩ refers to a set of problems; and ⟨1⟩t is the t-th point in ⟨1⟩.
III Predicting state.
The t-th cubature point’s predicted state from time k − 1 to time k is defined as

λ∗t,k|k−1 = g(λt,k−1|k−1 ). (9)

Then, the predicted state from time k − 1 to time k is obtained from (9),

2n
1
λ̂k|k−1 =
2n ∑ λ∗t,k|k−1 , (10)
t =1

and its covariance is

2n
1
Bk|k−1 =
2n ∑ λ∗t,k|k−1 λ∗t,kT|k−1 −
t =1 (11)
λ̂k|k−1 λ̂kT|k−1 + q k −1 .

2. The measurement update is as follows, including the error covariance Bk|k−1 at time k:
I Factorizing the CM of the error Bk|k−1 .

Bk|k−1 = Ak/k−1 AkT|k−1 . (12)

II Calculating the cubature points.

λt,k|k−1 = Ak|k−1 ζ t + λ̂k|k−1 . (13)

III Updating observation.

The estimated observation of the t-th cubature point between epochs k − 1 and k
is denoted by
∗
θt,k |k−1 = h ( λt,k|k−1 ). (14)
From (24), we can obtain the predicted observation of the t-th cubature point
from time k − 1 to k :
p 2n ∗
2n t∑
θ̂k|k−1 = θ k | k −1 , (15)
=1
and its covariance and cross-covariance matrices are
2n
p
Bθθ,k|k−1 =
2n ∑ θt,k
∗ ∗T
|k−1 θt,k|k−1 −
t =1 (16)
θ̂k|k−1 θ̂kT|k−1 + rk ,

and
2n
p
Bλθ,k|k−1 =
2n ∑ λ∗t,k|k−1 θt,k
∗T
| k −1 −
t =1 (17)
λ̂k|k−1 θ̂kT|k−1 .

IV Calculating the Kalman gain.

−1
Gk = Bλθ,k|k−1 Bθθ,k | k −1
. (18)
Sensors 2024, 24, 392 6 of 18

V State update.
λ̂k|k = λ̂k|k−1 + Gk (θk − θ̂k|k−1 ). (19)

VI CM of the estimate error update.

Bk|k = Bk|k−1 + Gk Bθθ,k|k−1 GkT . (20)

Figure 2 shows a comprehensive overview of the algorithm of the CKF with miss-
ing measurements.

Figure 2. The algorithm of the CKF with missing measurements.

4. Robust Cubature Kalman Filter with Missing Measurements

In the context of applying the CKF to a nonlinear system, it is imperative to possess
a comprehensive understanding of the noise statistics associated with the device as well
as its mathematical model. However, in the event that a filter is established based on an
inaccurate mathematical model and noise statistics, there is a possibility of encountering
a significant inaccuracy in the estimation of the system’s state or even the divergence of
the estimation [16]. The robust M-estimation theory is a valuable technique for estimating
unknown noise statistics [28]. Robust M-estimation can be employed to identify anomalies
in state estimation. Additionally, the continuous updating of the statistical features of
measurement noise enables the CKF to adapt to variations in the statistical characteristics
of measurement noise in real time. The RCKF technique is formed by integrating Huber’s
M-estimation theory with the conventional CKF model [26]. In this paper, this technique is
used to deal with nonlinear systems with missing measurements. The algorithm will be
Sensors 2024, 24, 392 7 of 18

derived in the subsequent sections. In contrast to the conventional CKF method, the RCKF
technique selectively modifies and updates the appropriate representations within the
measurement updating formula:

2n
p
B̃θθ,k|k−1 =
2n ∑ θt,k
∗ ∗T
|k−1 θt,k|k−1 − θ̂k|k−1 θ̂k|k−1 + r̃k , (21)
t =1

where B̃θθ,k|k−1 can be obtained by estimating a weight matrix B using an absence of

difference M-estimation approach, and r̃k is equal to the measurement noise variance
matrix rk . That is,
r̃k = B̃−1 , (22)
where the matrix B̃ is created using Huber’s approach [29]. This process depends on
considering the KF as a linear regression problem, as explained in [28], that can be solved
with resistance and robust efficiency using the M-estimation. This minimizes the cost
function as follows:
2n
′
C (λk ) = ∑ ρ ( bt ) . (23)
t =1
′
Here, bt denotes the residue vector’s t-th component
′
bt = bt /Abt , (24)

where bt is a residual component associated with the observation quantity θk , and Abt is
the mean square error associated with bt . The expressions Abt and bt are used in practice
because the covariance matrix of the measurement residuals is acquired from (16), which is
the variable quantity Bθθ,k|k−1 previous to being adjusted:

Abt = ( Bθθ,k|k−1 )tt , (25)

bt = (θk − θ̂k|k−1 )t . (26)

′
The score function ρ(bt ) is defined as follows [30,31]:
( ′ ′
′ bt2 if | bt |≤ c
ρ ( bt ) = ′ (27)
c | bk | − 12 c2 otherwise,

where c is a constant that is typically between 1.3 and 2.0 [16]. When the partial derivative
of (24) is set to zero,
2n ′ ′
∂ρ(bt ) ∂(bt )
∑ ′ . λt,k k = 1, 2, . . . , n, (28)
t = 1 ∂ ( bt )

where λt,k is the state vector in the t-th component. Following

′
∂ ( bt )
wt = ′ ′ , (29)
bt ∂bt

we can obtain the formula


′
1 if | bt |≤ c,
wt = c (30)
 ′ otherwise.
| bt |

Depending on (30), the Huber approach will determine which diagonal components
of B̃ are positive. An identical expression is provided below:
Sensors 2024, 24, 392 8 of 18

 ′
1 bt

At,t if | Abt |=| bt |≤ c
B̃t,t = (31)
 c ′ otherwise,
At,t |bt |

 ′ ′
 A1 if | bt |< c, | b j |< c
t,j
B̃t,j = c (32)
 ′ ′ otherwise,
At,j max(|bt |,|b j |)

where the diagonal and off-diagonal elements in the matrix B̃ are denoted as B̃t,t and B̃t,j ,
respectively. Similarly, the diagonal and off-diagonal elements in the measurement noise
rk are represented as At,t and At,j , respectively. The element At,j is equal to zero due to
the fact that the matrix representing the covariance of measurement noise is diagonal.
′
The symbol bt represents the measurement residual, while bt denotes the standard residual
error. Additionally, Abt represents the mean variance of bt . The algorithm for the given
RCKF with missing measurements is depicted in Figure 3.

Figure 3. The algorithm of the RCKF with missing measurements.

5. Metrics of Performance
When evaluating a new filter, it is commonly compared to standard filters using
benchmark workloads. The root mean square error (RMSE) is a widely used metric for
evaluation [32]. Still, it only assesses the filter’s output at the initial instance, specifically
the conditional mean of the state [33]. In this study, we not only compared the state
estimate λ̂k|k but also the conditional mean of the estimated error. The non-credibility index
(NCI) served as a credibility metric for comparing the filter’s efficiency in producing the
conditional mean [34].
1. Root mean square error (RMSE).
Sensors 2024, 24, 392 9 of 18

The RMSE for the state estimate λ̂k|k generated utilizing M Monte Carlo simulations
at time instant k is as follows:
v
u1 M
u

M t∑
RMSE(λ̂k|k ) = t (λk (t) − λ̂k|k )T (λk (t) − λ̂k|k ), (33)
=1

where the estimate of λk (t) at the t-th Monte Carlo simulation is λ̂k|k .
2. Non-credibility index.
In order to calculate the NCI, we compared the estimator’s normalized squared
estimation error, which is defined as

ϵk|k (t) = (λk (t) − λ̂k|k ) T Bk|k (t)−1 (λk (t) − λ̂k|k ), (34)

with the credible estimator’s normalized squared estimation error, expressed as [35]

ϵk∗|k (t) = (λk (t) − λ̂k|k ) T ϕk−|k1 (λk (t) − λ̂k|k ), (35)

1
where ϕk|k is the mean square error (MSE) computed by ( M ∑tM=1 (λk (t) − λ̂k|k )T (λk (t) −
λ̂k|k ). The NCI is described as

M ϵk | k ( t )
10
NCI (k) =
M ∑ log10
ϵk∗|k (t)
. (36)
t =1

The NCI can measure the estimator’s credibility. That is, the estimator’s CM is close
to the MSE (ϕk|k ). The lower the NCI score, the more reliable the estimator; therefore,
an NCI score of zero indicates an entirely credible estimator.

6. Numerical Experiments
This section presents a comparative analysis of the performance of the EKF, UKF,
CKF, and RCKF through the examination of two examples. The simulation of signal and
observation values was conducted using MATLAB, and alternate filtering estimates will be
presented. The determination of appropriate model parameters and how we conducted a
comprehensive study to compare the methods are also explained.

6.1. Model Specifications

This part describes the benchmark models used to compare the methods. The consis-
tently accelerating moving-target tracking model monitors a moving target with missing
measurements and has many applications. The numerical representations were carried out
using two different models: the UNGM [36] and BOT [37,38]. We applied the two models in
two scenarios regarding the missing measurements: p(ψk = 1) = 0.7 and p(ψk = 1) = 0.8.
1. The UNGM.
This model is described as follows:
λ k −1
λk =0.5λk−1 + 25 + 8cos(1.2(k − 1))
1 + λ2k−1 (37)
+ δk ,

and
λ2k
θk = ψk × + ξk, (38)
20
where δk ∼ N (0, 1), ξ k ∼ N (0, 1), λ0 ∼ N (0.1, 1), probability p(ψk = 1) = 0.7,
and p(ψk = 1) = 0.8.
2. BOT.
Sensors 2024, 24, 392 10 of 18

There are two states inside the bearing-only tracking (BOT) paradigm, with the state
λk = [λ1,k λ2,k ] T displaying a tracked target’s positioning in Cartesian coordinates.
Its nonlinear model is as follows:
 
0.9 0
λk = × λk−1 + δk , (39)
0 1

λ2,k − sin(k)
θk = ψk (arctan( )) + ξ k , (40)
λ1,k − cos(k)
   
1 0 T 1 0
where δk ∼ N (0, 0.001 × ), ξ k ∼ N (0, 0.005), λ0 ∼ N ([20 5] , 0.1 × ),
0 3 0 3
and p(ψk = 1) = 0.7, p(ψk = 1) = 0.8.

6.2. Experiment and Analysis

The UNGM and BOT examples underwent evaluation through 100 independent Monte
Carlo simulations with 50 time intervals and two scenarios regarding the missing measure-
ments: p(ψk = 1) = 0.7 and p(ψk = 1) = 0.8. Figures 4–6 depict the evolution of the RMSE
with time for the UNGM, BOT λ1 , and BOT λ2 , respectively. Tables 2 and 3 provide the
average RMSE (ARMSE) for the UNGM and BOT. The RCKF exhibited the highest accuracy
and achieved the lowest RMSE for the UNGM example in the two scenarios, as shown in
Figure 4. The RCKF consistently outperformed the other methods in filtering precision.
Also, Table 2 shows that the RCKF had the lowest ARMSE, with values of 3.27 and 1.60
when p(ψk = 1) = 0.7 and p(ψk = 1) = 0.8, respectively. For BOT state λ1 , both the RCKF
and CKF showed a comparable estimating accuracy and outperformed the EKF and EnKF
in the two scenarios. Figure 5 and Table 3 indicate that the RCKF and CKF had the same
ARMSE in the two scenarios, while in the state λ2 , the RCKF and CKF demonstrated similar
estimation accuracies when p(ψk = 1) = 0.7. The RCKF had the highest accuracy when
p(ψk = 1) = 0.8, achieving the lowest ARMSE of 0.40. Table 2 also reveals that the RCKF,
CKF, and EKF exhibited similar ARMSE values.
Figure 7 depicts the temporal progression of the NCI in the two scenarios for the
UNGM, while Figures 8 and 9 exhibit the corresponding evolution for BOT states λ1 and
λ2 in the two scenarios, respectively. Tables 4 and 5 present the average NCI (ANCI) for the
UNGM and BOT in both states. According to the findings in Figure 7, the RCKF exhibited
superior outcome credibility for the UNGM in the two scenarios, as seen by the lowest
scores. According to Table 4, the RCKF approach demonstrated outstanding performance
compared to the other methods regarding the average filtering credibility, as indicated by
the ANCI scores of 4.83 and 7.10 when p(ψk = 1) = 0.7 and p(ψk = 1) = 0.8, respectively.
Similarly, it can be observed from Figures 8 and 9 that the RCKF demonstrated the highest
level of reliability in the context of the BOT example. Additionally, it is worth noting that
the NCI score was found to be the lowest in both states. According to Table 5, the relative
credibility Kalman filter (RCKF) demonstrated exceptional performance in terms of average
filtering credibility in both states. Specifically, the ANCI values for λ1 and λ2 were reported
as 3.75 and 2.65, respectively, when p(ψk = 1) = 0.7, while the ANCI values were 0.18 and
0.29 when p(ψk = 1) = 0.8, indicating the superiority of the RCKF method compared to
other approaches in the two scenarios.
Sensors 2024, 24, 392 11 of 18

(a) p(ψk = 1) = 0.7

(b) p(ψk = 1) = 0.8

Figure 4. RMSE with time in the UNGM.

(a) p(ψk = 1) = 0.7

Figure 5. Cont.
Sensors 2024, 24, 392 12 of 18

(b) p(ψk = 1) = 0.8

Figure 5. RMSE with time in the BOT model for λ1 .

(a) p(ψk = 1) = 0.7

(b) p(ψk = 1) = 0.8

Figure 6. RMSE with time in the BOT model for λ2 .

Sensors 2024, 24, 392 13 of 18

Table 2. The UNGM average RMSE.

Average RMSE When Average RMSE When

Method
p(ψk = 1) = 0.7 p(ψk = 1) = 0.8
EKF 5.41 4.63
EnKF 4.82 5.32
CKF 3.60 4.16
RCKF 3.27 1.60

Table 3. BOT average RMSE.

Average RMSE λ1 Average RMSE λ1 Average RMSE λ2

Method When When Average RMSE λ2 When
p(ψk = 1) = 0.7 p(ψk = 1) = 0.8 p(ψk = 1) = 0.8
EKF 17.50 17.19 0.64 0.71
EnKF 74.38 63.78 22.79 20.20
CKF 0.067 0.07 0.55 0.54
RCKF 0.057 0.07 0.53 0.40

(a) p(ψk = 1) = 0.7

(b) p(ψk = 1) = 0.8

Figure 7. NCI with time in the UNGM.

Sensors 2024, 24, 392 14 of 18

(a) p(ψk = 1) = 0.7

(b) p(ψk = 1) = 0.8

Figure 8. NCI with time in the BOT model for λ1 .

(a) p(ψk = 1) = 0.7

Figure 9. Cont.
Sensors 2024, 24, 392 15 of 18

(b) p(ψk = 1) = 0.8

Figure 9. NCI with time in the BOT model for λ2 .

Table 4. The UNGM average NCI.

Average NCI When Average NCI When

Method
p(ψk = 1) = 0.7 p(ψk = 1) = 0.8
EKF 9.08 7.58
EnKF 9.44 11.07
CKF 5.23 7.46
RCKF 4.83 7.10

Table 5. BOT average NCI.

Average NCI λ1 Average NCI λ1 Average NCI λ2 Average NCI λ2

Method When When When When
p(ψk = 1) = 0.7 p(ψk = 1) = 0.8 p(ψk = 1) = 0.7 p(ψk = 1) = 0.8
EKF 3.91 0.69 6.91 6.76
EnKF 3.89 0.47 3.72 0.88
CKF 4.71 0.40 3.53 1.52
RCKF 3.75 0.18 2.65 0.29

7. Conclusions
This study presented the RCKF as a filter for nonlinear systems with missing mea-
surements. In order to accomplish this objective, we combined Huber’s M-estimation
theory with the conventional CKF for nonlinear systems with missing observations and
developed the filter using a recursive method. We demonstrated the effectiveness of the
proposed method through two examples, the UNGM and BOT, and compared it with
the EKF, EnKF, and CKF. The results showed that the RCKF provided more precise and
credible outcomes compared to the other methods, with the highest accuracy observed in
the UNGM example. Also, in the BOT example, the RCKF exhibited essentially superior
accuracy to other methods. In general, compared to traditional techniques such as the EKF,
EnKF, and CKF, the RCKF demonstrated the best accuracy and credibility for nonlinear
systems with missing measurements.
Future research will focus on extending the RCKF to capture missing measurement
phenomena through a general Markov chain rather than a Bernoulli sequence of identical
independent distributions. Additionally, we propose using the RCKF as an alternate
approach for estimating the state of nonlinear systems when the system noises follow a
non-Gaussian distribution instead of a Gaussian distribution.
Sensors 2024, 24, 392 16 of 18

Author Contributions: Conceptualization, E.S.; Methodology, S.S. and D.N.; Software, S.S.; Validation,
E.S.; Writing—original draft, S.S.; Writing—review & editing, E.S. and D.N.; Funding acquisition, E.S.
All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by National Natural Science Foundation of China grant number
U2066203, and Sichuan University grant number 2020SCUNG205.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Data are contained within the article.
Conflicts of Interest: The authors declare no conflict of interest.

Abbreviations
The following abbreviations are used in this manuscript:

Abbreviations
RCKF Robust cubature Kalman filter
CKF Cubature Kalman filter
EKF Extended Kalman filter
UKF Unscented Kalman filter
EnKF Ensemble Kalman filter
CM Covariance matrix
MSE Mean square error
RMSE Root mean square error
NCI Non-credibility index
UNGM Univariate Non-stationary Growth Model
BOT Bearing-only tracking
Symbols
λk The state vector n×1
θk The measurement vector m×1
g (.) Nonlinear function of the state n×n
h (.) Nonlinear function of the measurement m×m
δk−1 The process noise n×1
ξk The measurement noise m×1
q k −1 The covariance matrix of δk−1 n×n
rk The covariance matrix of ξ k m×m
ψk Factor of missing measurement 1×1
λ̂k|k−1 The predicted state estimation n×1
θ̂k|k−1 The predicted measurement estimation m×1
Bk|k−1 Predicted error covariance estimation n×n
Bθθ,k|k−1 Estimated matrix of innovation covariance m×m
Bλθ,k|k−1 Estimated cross-covariance matrix n×m
Gk Kalman gain n×m
λ̂k|k Estimated update state n×1
ζt The cubature point n×1
Estimated matrix of innovation covariance using
B̃θθ,k|k−1 m×m
an absence of difference M-estimation approach ξ k
bt Residue vectors t-th m×1
bt′ Residue vectors t-th m×1
Abt Mean variance of bt m×m

References
1. Kalman, R. A New Approach to Liner Filtering and Prediction Problems, Transaction of ASME. J. Basic Eng. 1961, 83, 95–108.
[CrossRef]
2. Radhakrishnan, R.; Bhaumik, S.; Tomar, N.K. Continuous-discrete filters for bearings-only underwater target tracking problems.
Asian J. Control 2019, 21, 1576–1586. [CrossRef]
Sensors 2024, 24, 392 17 of 18

3. Haus, B.; Mercorelli, P. An extended Kalman filter for time delays inspired by a fractional order model. In Proceedings of the
Conference on Non-integer Order Calculus and Its Applications, Lodz, Poland, 11–13 October 2017; Springer: Berlin/Heidelberg,
Germany, 2017; pp. 151–163.
4. Schimmack, M.; Haus, B.; Mercorelli, P. An extended Kalman filter as an observer in a control structure for health monitoring of a
metal–polymer hybrid soft actuator. IEEE/ASME Trans. Mechatron. 2018, 23, 1477–1487. [CrossRef]
5. Brossard, M.; Barrau, A.; Bonnabel, S. A code for unscented Kalman filtering on manifolds (UKF-M). In Proceedings of the 2020
IEEE International Conference on Robotics and Automation (ICRA), Paris, France, 31 May– 31 August 2020; IEEE: Piscataway, NJ,
USA, 2020; pp. 5701–5708.
6. Gao, B.; Gao, S.; Hu, G.; Zhong, Y.; Gu, C. Maximum likelihood principle and moving horizon estimation based adaptive
unscented Kalman filter. Aerosp. Sci. Technol. 2018, 73, 184–196. [CrossRef]
7. Nilam, B.; Ram, S.T. Forecasting Geomagnetic activity (Dst Index) using the ensemble kalman filter. Mon. Not. R. Astron. Soc.
2022, 511, 723–731. [CrossRef]
8. Gillijns, S.; Mendoza, O.B.; Chandrasekar, J.; De Moor, B.; Bernstein, D.; Ridley, A. What is the ensemble Kalman filter and how
well does it work? In Proceedings of the 2006 American Control Conference, Minneapolis, MN, USA, 14–16 June 2006; IEEE:
Piscataway, NJ, USA, 2006; p. 6.
9. Katzfuss, M.; Stroud, J.R.; Wikle, C.K. Understanding the ensemble Kalman filter. Am. Stat. 2016, 70, 350–357. [CrossRef]
10. Zhou, N.; Meng, D.; Huang, Z.; Welch, G. Dynamic state estimation of a synchronous machine using PMU data: A comparative
study. IEEE Trans. Smart Grid 2014, 6, 450–460. [CrossRef]
11. Arasaratnam, I.; Haykin, S. Cubature Kalman filters. IEEE Trans. Autom. Control 2009, 54, 1254–1269. [CrossRef]
12. Mallick, M.; Tian, X.; Liu, J. Evaluation of measurement converted KF, EKF, UKF, CKF, and PF in GMTI filtering. In Proceedings
of the 2021 International Conference on Control, Automation and Information Sciences (ICCAIS), Xi’an, China, 14–17 October
2021; IEEE: Piscataway, NJ, USA, 2021; pp. 21–27.
13. Liu, W.; Zhang, Q.; Zhao, Y. Performance Evaluation of Cubature Kalman Filtering and Extended Kalman Filtering Based Phase
Unwrapping For Insar. In Proceedings of the IOP Conference Series: Earth and Environmental Science, Surakarta, Indonesia,
24–25 August 2021; IOP Publishing: Bristol, UK, 2021; Volume 693, p. 012047.
14. Zhao, L.; Wang, R.; Wang, J.; Yu, T.; Su, A. Nonlinear state estimation with delayed measurements using data fusion technique
and cubature Kalman filter for chemical processes. Chem. Eng. Res. Des. 2019, 141, 502–515. [CrossRef]
15. Zhang, Y.; Wang, J.; Sun, Q.; Gao, W. Adaptive cubature Kalman filter based on the variance-covariance components estimation.
J. Glob. Position. Syst. 2017, 15, 1–9. [CrossRef]
16. Zhao, X.; Li, J.; Yan, X.; Ji, S. Robust adaptive cubature Kalman filter and its application to ultra-tightly coupled SINS/GPS
navigation system. Sensors 2018, 18, 2352. [CrossRef] [PubMed]
17. Zhang, C.; Zhi, R.; Li, T.; Corchado, J. Adaptive m-estimation for robust cubature kalman filtering. In Proceedings of the 2016
Sensor Signal Processing for Defence (SSPD), Edinburgh, UK, 22–23 September 2016; IEEE: Piscataway, NJ, USA, 2016; pp. 1–5.
18. Cui, B.; Chen, X.; Tang, X.; Huang, H.; Liu, X. Robust cubature Kalman filter for GNSS/INS with missing observations and
colored measurement noise. ISA Trans. 2018, 72, 138–146. [CrossRef] [PubMed]
19. Ye, X.; Wang, J.; Wu, D.; Zhang, Y.; Li, B. A Novel Adaptive Robust Cubature Kalman Filter for Maneuvering Target Tracking
with Model Uncertainty and Abnormal Measurement Noises. Sensors 2023, 23, 6966. [CrossRef] [PubMed]
20. Atitallah, M.; Davoodi, M.; Meskin, N. Event-triggered fault detection for networked control systems subject to packet dropout.
Asian J. Control 2018, 20, 2195–2206. [CrossRef]
21. Liu, W.Q.; Tao, G.L.; Fan, Y.J.; Zhang, G.Q. Robust fusion steady-state filtering for multisensor networked systems with one-step
random delay, missing measurements, and uncertain-variance multiplicative and additive white noises. Int. J. Robust Nonlinear
Control 2019, 29, 4716–4754. [CrossRef]
22. Xu, L.; Ma, K.; Fan, H. Unscented Kalman filtering for nonlinear state estimation with correlated noises and missing measurements.
Int. J. Control Autom. Syst. 2018, 16, 1011–1020. [CrossRef]
23. Zhang, X.; Yan, Z.; Chen, Y. High-degree cubature Kalman filter for nonlinear state estimation with missing measurements. Asian
J. Control 2022, 24, 1261–1272. [CrossRef]
24. Fang, H.; Tian, N.; Wang, Y.; Zhou, M.; Haile, M.A. Nonlinear Bayesian estimation: From Kalman filtering to a broader horizon.
IEEE/CAA J. Autom. Sin. 2018, 5, 401–417. [CrossRef]
25. Meng, Q.; Leib, H.; Li, X. Cubature ensemble Kalman filter for highly dimensional strongly nonlinear systems. IEEE Access 2020,
8, 144892–144907. [CrossRef]
26. Li, Y.; Li, J.; Qi, J.; Chen, L. Robust cubature Kalman filter for dynamic state estimation of synchronous machines under unknown
measurement noise statistics. IEEE Access 2019, 7, 29139–29148. [CrossRef]
27. Pu, Y.; Li, X.; Liu, Y.; Wang, Y.; Wu, S.; Qu, T.; Xi, J. Improved Strong Tracking Cubature Kalman Filter for UWB Positioning.
Sensors 2023, 23, 7463. [CrossRef] [PubMed]
28. Durovic, Z.M.; Kovacevic, B.D. Robust estimation with unknown noise statistics. IEEE Trans. Autom. Control 1999, 44, 1292–1296.
[CrossRef]
29. Gui, Q.; Zhang, J. Robust biased estimation and its applications in geodetic adjustments. J. Geod. 1998, 72, 430–435. [CrossRef]
30. Huang, Y.; Wu, L.; Sun, F. Robust Cubature Kalman filter based on Huber M estimator. Control Decis. 2014, 29, 572–576.
Sensors 2024, 24, 392 18 of 18

31. Li, K.; Hu, B.; Chang, L.; Li, Y. Robust square-root cubature Kalman filter based on Huber’s M-estimation methodology. Proc.
Inst. Mech. Eng. Part G J. Aerosp. Eng. 2015, 229, 1236–1245. [CrossRef]
32. Hodson, T.O. Root-mean-square error (RMSE) or mean absolute error (MAE): When to use them or not. Geosci. Model Dev. 2022,
15, 5481–5487. [CrossRef]
33. Li, X.R.; Zhao, Z. Measuring estimator’s credibility: Noncredibility index. In Proceedings of the 2006 9th International Conference
on Information Fusion, Florence, Italy, 10–13 July 2006; IEEE: Piscataway, NJ, USA, 2006; pp. 1–8.
34. Rohr, D.; Lawrance, N.; Andersson, O.; Siegwart, R. Credible Online Dynamics Learning for Hybrid UAVs. In Proceedings of
the 2023 IEEE International Conference on Robotics and Automation (ICRA), London, UK, 29 May–2 June 2023; pp. 1305–1311.
[CrossRef]
35. Bar-Shalom, Y.; Li, X.R.; Kirubarajan, T. Estimation with Applications to Tracking and Navigation: Theory Algorithms and Software; John
Wiley & Sons: Hoboken, NJ, USA, 2004.
36. Tong, Y.; Zheng, Z.; Fan, W.; Li, Q.; Liu, Z. An improved unscented Kalman filter for nonlinear systems with one-step randomly
delayed measurement and unknown latency probability. Digit. Signal Process. 2022, 121, 103324. [CrossRef]
37. Julier, S.J.; Uhlmann, J.K. New extension of the Kalman filter to nonlinear systems. In Proceedings of the Signal Processing,
Sensor Fusion, and Target Recognition VI, Orlando, FL, USA, 21–24 April 1997; SPIE: Bellingham, WA, USA, 1997; Volume 3068,
pp. 182–193.
38. Kotecha, J.H.; Djuric, P.M. Gaussian sum particle filtering. IEEE Trans. Signal Process. 2003, 51, 2602–2612. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.