I.

INTRODUCTION
Multiple target tracking [1], [2] includes state estimation
of the targets of interest from a set of noisy measurements
A Learning Gaussian Process and false alarms. Multiple target tracking is part of vari-
ous automation systems in diverse fields. The application
Approach for Maneuvering spectrum ranges from micro level, such as human cell
trackers [3], to macro level, e.g., tracking of aircrafts [1],
Target Tracking and Smoothing [2], [4], [5]. Other applications include oceanography [6],
sea surveillance systems [7], autonomous vehicles [8], area
security [9], and many more.
Multiple target tracking algorithms are sometimes clas-
sified as point, extended, and group target tracking [10],
WAQAS AFTAB [11]. The sensor generates multiple measurements for an ex-
LYUDMILA MIHAYLOVA , Senior Member, IEEE tended target. Comparatively, scarce (in some cases single)
The University of Sheffield, Sheffield, measurements for a point target are assumed. The process
of tracking of multiple point targets possessing similar
kinematics, can be performed as a one group and is called
group target tracking [10]. The point target tracking requires
Model-based approaches for target tracking and smoothing es- estimation of the target kinematics such as position and
timate the infinite number of possible target trajectories using a velocity, attributes and other parameters [5]. The extended
finite set of models. This article proposes a data-driven approach
that represents the possible target trajectories using a distribution
target tracking and group target tracking, in addition to the
over an infinite number of functions. Recursive Gaussian process, and abovementioned parameters of the point objects, estimate
derivative-based Gaussian process approaches for target tracking, the target shape, volume, orientation, and other similar
and smoothing are developed, with online training, and parameter parameters. The approach presented in this article relates
learning. The performance evaluation over two highly maneuvering to point target tracking.
scenarios, shows that the proposed approach provides 80 and 62%
performance improvement in the position, and 49 and 22% in the
Two major processes involved in point target tracking
velocity estimation, respectively, as compared to the best model-based are the data association, i.e., measurement to target/track
filter. assignment, and the state estimation, which includes tar-
get state update using the assigned measurement [5]. The
performance of these two processes is interdependent. The
output of the data association is considered as an input
to the state estimation at the same time step. The output
of the state estimation is an input to the data associa-
tion process of the subsequent time step. Many multiple
target tracking approaches solve the data association and
the state estimation problems independently such as the
global nearest neighbor tracker [1], [2], [5], multiple hy-
Manuscript received December 16, 2019; revised April 7, 2020 and July
pothesis tracker [12], and probabilistic data association
20, 2020; released for publication July 21, 2020. Date of publication filter tracker [1]. Other approaches provide a simultaneous
September 2, 2020; date of current version February 9, 2021. solution, e.g., the multiscan Markov chain Monte Carlo data
DOI. No. 10.1109/TAES.2020.3021220
association tracker [13] and the random finite sets based
approaches, such as the probability hypothesis density filter
Refereeing of this contribution was handled by G. Hendeby. trackers [14]. In all the multiple target tracking approaches,
The work of W. Aftab was supported by the Government of Pakistan and the state estimation process can be treated as an independent
Department of ACSE (University of Sheffield). This work was supported process. In this article, the measurement to track assignment
in part by the U.K. Engineering and Physical Sciences Research Council is assumed known and the solution to the state estimation
through EP/T013265/1 Project NSF-EPSRC:ShiRAS Towards Safe and
Reliable Autonomy in Sensor Driven Systems and in part by the EPSRC
problem is provided.
for the support through the Bayesian Tracking and Reasoning over Time Typically, the estimation problem is formulated within
(BTaRoT) under Grant EP/K021516/1. This work was also supported by a Bayesian framework. The target dynamics is considered
the USA National Science Foundation under Grant NSF ECCS 1903466. nonlinear in most applications. Sometimes, the measure-
Authors’ address: The authors are with the Department of Auto- ment process is also considered nonlinear. Different filtering
matic Control and Systems Engineering, The University of Sheffield, methods [1], [5] have been proposed which depend on the
Sheffield S10 2TG, U.K., E-mail: ([email protected]; extent of model nonlinearity, the required accuracy and the
[email protected]). (Corresponding author: Waqas Aftab.) available processing time. The focus of this article is in
real-time estimation only and for the first time, an online
This work is licensed under a Creative Commons Attribution 4.0 License. data driven approach for point target tracking and fixed-lag
For more information, see https://creativecommons.org/licenses/by/4.0/ smoothing is proposed.

278 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 1 FEBRUARY 2021
A. Related Work in [28], provides a solution to the data association problem
For linear target dynamics and measurement models, the of point targets. To the best knowledge of the authors, this is
well-known Kalman filter (KF) [15] provides a recursive the first work on a GP-based model-free approach for online
optimal solution. For nonlinear models, many filters have target kinematics filtering and smoothing.
been proposed such as an extended KF [16], an unscented B. Contributions
KF [17], and an interacting multiple model (IMM) [18]
filter. The IMM has been shown to be one of the most cost The key contributions of this article are as follows.
effective nonlinear filters [19]. The KFs based on different 1) A data-driven approach is proposed for online point
nonlinear target dynamics models have also been proposed target tracking. The proposed approach is shown
for tracking nonlinear kinematics, such as the coordinated to estimate well the unknown and nonlinear target
turn [2] KF and Singer [20] Kalman Filter (KF). All of the trajectory using the recursive update of the GP hy-
above approaches rely on an underlying target dynamics perparameters (see Section III-B).
model. A model, closest to the average target trajectory, 2) A model-free recursive approach is proposed for the
is chosen and the variations are captured using the model estimation of the derivative of a GP. This approach
noise. A highly maneuverable target such as an aircraft can also be extended in the same way to recursively
can adopt an infinite number of trajectories. Additionally, update the higher order derivatives of the GP. The
the model variations are expected to be large. In such proposed approach is demonstrated by estimating the
cases, the single model-based filters require a high model velocity of a point target (see Section III-C).
noise variance to track the different trajectories. A high 3) A recursive point target smoother is proposed for
noise variance degrades the estimation performance and online position and velocity smoothing (see Sec-
gives inaccurate estimates even when the target trajectory tion III-D). The proposed smoother can be extended
matches the model. The multiple model-based filters are for smoothing of the higher order derivatives of the
an extension of the single model-based filters. These are GP in the same way.
again limited in performance as all the possible trajectories 4) The proposed filters and smoothers are shown to be
are not modeled exactly. A preliminary study on a non- robust to the measurement noise model parameters.
recursive model-free filtering, a Gaussian process motion This is achieved by augmenting the state vector
tracker (GPMT), was proposed in [21]. This is a Gaussian with the measurement noise variance parameter. This
process (GP)-based approach, which predicts and filters parameter is recursively updated at each time sam-
the position and higher order position derivatives using ple (see Section III-E). A simulation-based study
historical measurement data. The Gaussian Process Motion is also carried out to demonstrate the robustness of
Tracker (GPMT), includes training and learning stages, the proposed approaches to the measurement noise
models infinite number of target trajectories. The learning variance changes (see Section IV-F).
stage proposed for the GPMT, is not recursive. In this article, 5) The performance evaluation of the proposed ap-
we propose a Recursive Gaussian Process Motion Tracker proach is done under challenging scenarios and
(RGP MT)1 for online model-free point target tracking. The compared with model-based approaches (see Sec-
state estimation and learning of the GP hyperparameters tion IV).
are performed recursively. The historical data is replaced
with an inducing point set representation for the recursive The remaining part of this article is structured as follows.
filtering and smoothing. Background knowledge for GP and recursive GP methods
The GP regression method has been applied to many is given in Section II. The proposed tracker and smoother
areas involving time series prediction [22], [23]. Target state are introduced in Section III. The performance evaluation
filtering and smoothing is also a time series application but and results are presented in Section IV. Finally, Section V
the Gaussian Process (GP)-based methods have not been concludes this article.
studied widely by the target tracking community. GP meth-
ods have been proposed for localization [24], [25]. However, II. BACKGROUND KNOWLEDGE
the target trajectory in these works is not considered highly A. GP Regression
maneuverable. This article focuses on both highly maneu-
verable and less agile targets. Some GP-based methods have A GP is a stochastic process defining a distribution over
been proposed for extended object tracking [26], [27]. These possible functions that fit a set of points. A GP is a nonpara-
methods represent the object shape using a GP, whereas metric method [23] which can have different probability
the target kinematics estimation is performed using model- distributions to fit data to the unknown functions. Most
based approaches for point target tracking. The approach commonly, the squared exponential and Gaussian kernels
proposed in this article is for the kinematics estimation are used. The GP is defined by two functions, a mean and a
of point targets. An overlapping mixture of GP, proposed covariance kernel. It is assumed that any finite realizations
of the GP are mutually independent normally distributed
with the given mean and covariance kernel. The parameters
1 The indicates that the RGP MT is a recursive algorithm, whose hyper- of the mean and the covariance kernel are called hyper-
parameters are learnt online. parameters. In a regression problem, the function/model

AFTAB AND MIHAYLOVA: LEARNING GAUSSIAN PROCESS APPROACH FOR MANEUVERING TARGET TRACKING 279
 ⎡ ⎤
selection is done using the hyperparameters and the given ∂k (u,u1 ) ∂k (u,ul )
···
data also called training data. The hyperparameters are ⎢ ∂u u=u1 ∂u u=u1 ⎥
⎢ .. .. ⎥
determined using the training data and the process is called u (K uu ) = ⎢
⎢ .
..
. . ⎥
⎥ (7)
learning. ⎣ ∂k (u,u ) ∂k (u,ul )
⎦
Suppose a one-dimensional (1-D) output r is nonlinearly ∂u
1
··· ∂u
u=u j u=u j
dependent on a 1-D input u. The nonlinear mapping and the ⎡ ⎤
∂ 2 k (u,u ) ∂ 2 k (u,u )
observation model are given as ··· u=u1
⎢ ∂u∂u uu==uu1
 ∂u∂u
⎢ 1 u =ul ⎥
⎥
⎢ .. .. .. ⎥
r = f ( u ), f (u ) ∼ GP (m (u ), k (u, u )) (1) 2uu(K uu)= ⎢ . . . ⎥ (8)
⎢ ⎥
⎣ ∂ 2 k (u,u ) ∂ 2 k (u,ul ) ⎦
z = f (u ) + ,  ∼ N (0, σ I ) 2
(2) ∂u∂u u=u j ··· ∂ 2 ˜u∂u u=u j
u =u1 u =ul

where f is a nonlinear function, u and u are either the train-

ing or the testing input data, GP (m (u ), k (u, u )) represents where δm(u ) denotes the partial derivative of the mean func-
a GP with a mean m and a covariance kernel k, z is the tion, u (K uu ) and 2uu (K uu ) represent, respectively, the
observation vector, f (u ) represents the true function values first- and second-order derivative of the covariance matrix,
at input vector u,  denotes the measurement noise vector and · denotes the variable that is evaluated at α.
whose elements are independent and identically distributed α
(i.i.d.) Gaussian random variables with variance σ 2 and
I denotes an identity matrix. The n-dimensional identity B. Recursive GP Regression
matrix is written as I n . However, often we will not indicate
its dimension. GP regression is a powerful estimation method with
Given an n-dimensional training data set, Dn = high-computational training and learning stages. The com-
{(z1 , u1 ), (z2 , u2 ), . . . , (zn , un )}, the GP regression equa- putations scale as O (n3 ) for the n−dimensional training
tions at new input locations are data. The learning requires solving a nonconvex optimiza-
tion problem. Typically, an analytical solution to this opti-
mization problem is not available. Most of the existing nu-
E[ f (u )] = m(u )+K u u (K u u + σ 2 I n )−1 (z − m(u )) (3)
merical solutions are not applicable in real time. In [29], an
C[ f (u )] = K u u − K u u (K u u + σ 2 I n )−1 K uu (4) online approach for GP regression has been proposed. Two
methods have been proposed in [29] by assuming bounded
where E[·] denotes the mathematical expectation op- input domain. The first one addresses the computational
erator, f (u ) = [ f (u1 ), f (u2 ), . . . , f (ul )]T represents the complexity of the training process and is termed as online
function prediction vector at the new input vector u = regression or recursive GP (RGP). The second method, in
[u1 , . . . , ul ]T , E[ f (u )] and C[ f (u )] represent, respec- addition to the training, provides online learning and is
tively, the predicted mean and the predicted covariance, called online learning or recursive GP (RGP ). The input
z = [z1 , z2 , . . . , zn ]T and u = [u1 , u2 , . . . , un ]T are the n- domain is sampled at discrete points and the input vector
dimensional measurement and input training data vec- representing the sparse grid points is called basis vector u.
tors, respectively, m(u ) = [m (u1 ), m (u2 ), . . . , m (un )]T rep- The mean and the covariance at the sparse grid points form
resents the n-dimensional GP mean vector, K pq denotes the the sparse GP representation. The sparse GP representation
GP covariance matrix between the input vectors p and q, ·−1 is updated using the measurements received at each time
represents the matrix inverse. A GP covariance matrix sample.
between a j-dimensional vector u and the l-dimensional
vector u is given as
C. RGP With Online Regression
⎡ ⎤
k (u1 , u1 ) k (u1 , u2 ) · · · k (u1 , ul ) This section summarizes the RGP with online regression
⎢ ⎥ proposed in [29]. Consider the GP model of (1) and (2). The
⎢k (u2 , u1 ) k (u2 , u2 ) · · · k (u2 , ul )⎥
K uu = ⎢
⎢ .. .. .. .. ⎥.
⎥ (5) unknown function f is represented by a multivariate Gaus-
⎣ . . . . ⎦ sian distributed vector of N inducing points f = f (u ) =
k (u j , u1 ) k (u j , u2 ) · · · k (u j , ul ) [ f (u1 ), . . . , f (uN )]T with an initial distribution p0 ( f ) =
N (μ0f , C 0f ). At each time sample, the inducing point dis-
The above GP formulation is applicable to multiple dimen- tribution is updated using the corresponding measurements
sional inputs and outputs in a similar way. and the prior distribution, assuming known hyperparame-
The derivatives of the mean function and covariance ters. The required conditional distribution p( f |z1:k ) cannot
kernel are required and are defined as be updated using the GP regression (3) and (4). First,
because, the GP regression is a prediction method whereas
 T a measurement update is required. Second, the GP regres-
∂m (u ) ∂m (u )
δ(m(u )) = ∂u
··· ∂u (6) sion requires complete training data whereas a recursive
u=u1 u=u j update of a prior distribution is required. The conditional

280 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 1 FEBRUARY 2021
distribution is expanded as follows for the recursive update: online within the state space framework
⎡ ⎤
f k−1  
p( f |z1:k ) = c · p(zk | f , f̄ ) · p( f̄ | f ) · p( f |z1:k−1 )d f̄ ⎢ ⎥ f k−1
⎣ θ k−1 ⎦ = Ak + νk (21)
(9) θ k−1
f̄ k
where f̄ = f (ukf̄ ) denotes the functional evaluation at the yk = Ak xk−1 + νk (22)
⎡ ⎤
measurement input vector ukf̄
and c is a normalization I O
⎢ ⎥
constant. The product p( f̄ | f ) · p( f |z1:k−1 ) represents the Ak = ⎣ O I ⎦ , νk ∼ N (μνk , C νk ) (23)
predictive distribution. The RGP recursion equations are Jk O
Prediction : where xk denotes the state vector, yk is an augmented vector
Jk = K u f̄ u K −
uu
1
(10) composed of the state vector and the function prediction
k
at the measured locations, θ k = [ηk , σk ]T represents the
Bk = K u f̄ u f̄ − J k K uu f̄ (11) hyperparameters vector, ηk denotes the hyperparameters
k k k
vector of the GP mean and covariance functions, f and
μ̃kf̄ = m(ukf̄ ) + J k μ̂k−
f
1 − m (u ) (12) f̄ are, respectively, the function evaluation at the inducing
f̄ f points and measured locations, Ak is the state update matrix,
C̃ k = Bk + J k Ĉ k−1 J Tk (13) O denotes a zero vector/matrix of appropriate dimensions,
Estimation νk is the additive Gaussian noise vector with independent
and identically distributed (i.i.d.) elements and with the
M k = C̃ k (C̃ k + σ 2 I )−1
f̄ f̄
(14)
following parameters:
μ̂kf̄ = μ̃kf̄ + M k zk − μ̃kf̄ (15) ⎡ ⎤ ⎡ ⎤
O O O O
⎢ ⎥ ⎢ ⎥
f̄ f̄
Ĉ k = C̃ k − M k C̃ k
f̄
(16) μνk = ⎣ O ⎦ , C νk = ⎣O O O ⎦ (24)
f f̄ −1 f̄ bk O O Bk
C kf f̄ = Ĉ k−1 J Tk C̃ k Ĉ k (17)
f f̄ −1 where bk = m(ukf̄ ) − J k m(u ) and J k and Bk are given
Gk = Ĉ k−1 J Tk C̃ k +σ I 2
(18) by (10) and (11), respectively. The RGP recursion begins
with s sigma points selection around the mean hyperparam-
μ̂kf = μ̂k−
f f̄
1 + Gk zk − μ̃k (19) eters vector μθk−1 . In [29], the unscented transform [30] is
f f f used for sigma points selection. Given an ith sigma point θ ik
Ĉ k = Ĉ k−1 − Gk J k Ĉ k−1 (20)
with weight wik , the predicted state vector and covariance
matrix corresponding to the ith sigma point are in the form
where u denotes the input basis vector, ukf̄ represents the 
input vector corresponding to the kth measurement vector f  i θ

μk− + S θ − μ
+ μνk i
yi k
zk , J k and Bk are matrices derived from the GP regression (3) μk = Ak i 1 k k−1
(25)
θk θ ik θk
and (4), respectively, ˜˙ and ˆ˙ represent, respectively, the pre-  
dicted and filtered variables, μ f and C f are, respectively, f θf
C k− 1 − Sk C k−1 O
AT + C ν
yi
the sparse GP mean and covariance of the modeled non- C k = Ak i (26)
θk O O k θik k θik
linear function f at u, Gk is the gain matrix, σ 2 is the
f θ  θ −1
measurement noise variance hyperparameter, μ f̄ and C f̄ Sk = C k− 1 C k−1 (27)
represent, respectively, the estimated mean and covariance y y
of the unknown function evaluated at the measured location where μki and C ki represent, respectively, the state mean and
vector ukf̄ , C kf f̄ is the cross-covariance between f and f̄ , and the state error covariance corresponding to the ith sigma
M k is the Kalman gain matrix. point, C θk−1 represents the covariance of the hyperparam-
To summarize, the inducing points distribution at k − 1 eters vector and C θk−f 1 and C k−
fθ
1 are the cross-covariance
is used to find the predictive distribution of f̄ in (10)–(13). matrices between the hyperparameters vector and the sparse
The prediction step utilizes the GP regression of (3) and (4). GP representation f . The combined prediction variables are
The predicted distribution is updated using the measurement calculated next, as follows:
vector zk to give the posterior distribution using a KF [15] 
s
y
in (14)–(16). The product of Gaussians is evaluated for the μyk = wik μki (28)
estimated distribution of the inducing points in (18)–(20). i=1
s
y y y
D. RGP With Online Learning C yk = y y
wik ((μki − μk )(μki − μk )T + C ki ) (29)
i=1
This section summarizes the RGP with online learning
proposed in [29]. Consider the GP described by (1) and (2). where μyk and C yk denote, respectively, the mean and co-
The hyperparameters are assumed unknown and learned variance of the augmented vector y. The state vector xk

AFTAB AND MIHAYLOVA: LEARNING GAUSSIAN PROCESS APPROACH FOR MANEUVERING TARGET TRACKING 281
is decomposed in to an observed, ok = [σk , ( f̄ k )T ]T , and are updated as given below
unobserved vector, uk = [( f k )T , ηTk ]T . The mean vector and    
μ u
μkf
covariance matrix corresponding to this decomposition are μxk = k
= (40)
hT μ̂ok μθk
⎡ f ⎤  
  μk−1 o
⎢ η ⎥
C uk Lk Ĉ k h
μk−1
u
⎢μk−1 ⎥ Ck =
x
o o (41)
μyk = = ⎢ ⎥ (30) hT Ĉ k LTk hT Ĉ k h
μ̃ok ⎣μσk−1 ⎦
μ̃kf̄ where hT = [1, ONk ], ONk represents a 1 × Nk -dimensional
⎡ ⎤ zero vector, and Nk is the number of measurements received
f fη fσ f f̄
C k− C k− C k− C̃ k at k.
  ⎢ 1 1 1 ⎥
uo
C uk−1C̃ k ⎢C η f η ησ η f̄ ⎥
y ⎢ k−1 C k−1 C k−1 C̃ k ⎥
Ck = ou o = ⎢ σ f̄ ⎥
(31) III. PROPOSED DATA-DRIVEN RECURSIVE TRACKING
C̃ k C̃ k ⎢C σ f C σ η C σ ⎥ APPROACH
⎣ k−1 k−1 k−1 C̃ k ⎦
f̄ f f̄ η f̄ σ f̄
C̃ k C̃ k C̃ k C̃ k A GPMT was proposed by us in [21]. The GPMT deals
with 2-D (x and y) data. It can be extended to any number
where μ and μ represent, respectively, the mean of a vector of dimensions in a similar way. The GPMT assumes that
and a scalar, C a is the covariance matrix of vector a, C ab is the target motion, in the x and y directions, is mutually
the cross-covariance between vectors a and b, C ab and C ba uncorrelated. However, both coordinates are correlated in
denotes the covariance between scalar a and vector b and time. Furthermore, it is assumed that the motion correlation
C a is the variance of the scalar variable a. The measurement in time is limited to the recent past and is uncorrelated
model (2) is reformulated in the form with distant past. The filter has a sliding window of data
and as a result, the computational complexity of the GP
zkm = f (ukm ) + σ βkm , β ∼ N (0, 1) (32) training is reduced for online processing. In target tracking,
the higher order position derivatives are also estimated. The
to make σ explicit. In (32), β is assumed to be a random first-order GPMT (FO-GPMT2 ) was proposed in [21] for
variable sampled from the standard Gaussian distribution the velocity estimation. The same approach can be extended
(which is with a zero mean and covariance equal to 1) scaled for estimating other higher order derivatives. Importance of
by σ , zkm and ukm are, respectively, the mth measurement of Learning. A fixed set of hyperparameters allows tracking
zk and its corresponding input location, σ is the noise vari- of a limited set of target trajectories. The hyperparameters
ance (hyperparameter). The β noise is assumed uncorrelated need to be learned for tracking a relatively wide set of target
with σ . The mean and the covariances corresponding to the trajectories. In [21], a maximum likelihood (ML) based
model (32) are given as approach was adopted for learning. The computational cost
of this approach [21] is high and the GPMT processing is not
μzk = μ̃kf̄ (33) in realtime. A recursive GPMT (RGP MT) is proposed in
f̄ f̄ σ σ this article to reduce the computational time of the learning
C zk = C̃ k + V[σ ] + E[σ ]2 = C̃ k + Ck− 1 + (μk−1 )
2
(34) process. As a result, the RGP MT provides realtime target
k = C[ok , zk ] = C[ok , f̄ ]
C oz (35) state estimation.
The GPMT is summarized next followed by the pro-
where C[·] and V[·] represent, respectively, a covariance posed approach.
and variance function, μzk and C zk denote, respectively, the
mean and covariance of the predicted measurement vector A. GP Motion Tracker
and C oz
k is the cross-covariance between the observed state The GPMT is a GP-based approach for position predic-
and the predicted measurement. It is assumed that o and f̄ tion and filtering of nonlinear target dynamics. It is based
are jointly Gaussian. The KF equations used to update the on the following assumptions.
observed and unobserved states are
1) Cross-coordinates coupling is weak enough to be
μ̂ok = μ̃ok + E k (zk − μzk ) (36) ignored.
o o 2) Coordinate autocorrelation is available in time.
Ĉ k = C̃ k − E k C zk E Tk (37)
3) The temporal correlation with input points in distant
μuk = μuk−1 + Lk (μ̂ok − μ̃ok ) (38) past is weak.
o o 4) The measurement noise variance is the same for all
C uk = C uk−1 + Lk (Ĉ k − C̃ k )LTk (39)
the measurements.
z −1 uo o
where E k = C oz
k (C k ) and Lk = C̃ k (C̃ k )−1 are the Kalman The cross-coordinate coupling is ignored according to 1.
gain matrices for the observed and unobserved state updates, The coordinates are coupled when a target maneuvers [31].
uo
respectively, and C̃ k is the predicted cross-covariance ma-
trix between the unobserved and observed states and is a 2 The
submatrix of C yk . The state vector and the covariance matrix prefix FO is removed from hereon for brevity.

282 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 1 FEBRUARY 2021
The effect of ignoring this coupling in a model-free ap-
proach such as GPMT has not been studied yet. The co-
ordinate coupling can be included by extending the pro-
posed approach using a coupled GP [32]. The GPMT can
estimate any number of output dimensions. For tracking
of x and y coordinates, a 2D-GPMT is proposed in [21]
and summarized in this section. The GPMT formulation for
x-coordinate is presented below, since tracking of the y and
other coordinates is done in the same way. The x-coordinate
is represented using a GP Fig. 1. This figure illustrates the GPMT, RGP MT, and Recursive
Gaussian Process Motion Smoother (RGP MS) approaches for output at
x = f x (t ), f x ∼ GPx (0, k x (t, t  )) (42) k = 9 using a sliding window width d = 6. The target trajectory (solid
red line) is sampled at 8 points and the prediction is done for k = 9. The
ztx = f (t ) +
x
tx , tx ∼ N (0, σx2 I ) (43) GPMT uses d = 6 measurement samples (black dots), 3 ≤ k ≤ 8, as
training data. It predicts the trajectory at 9th sample shown as black
where x represents the horizontal Cartesian coordinate or empty circle. The RGP MT uses d = 6 inducing points (blue encircled
the output (position) variables, f x denotes the unknown plus) to predict the trajectory shown as blue plus. The RGP MT filtering
nonlinear function, GP denotes the GP representation, ztx at k = 9, updates the inducing points based on the new measurement
represents the measurement vector consisting of samples sample. The RGP MS estimates or the smoothed estimate (blue cross)
are the inducing points that are no more part of the training data. After
corresponding to time vector t, tx denotes an additive Gaus-
processing of the 9th sample, the training data at k = 3 is removed and
sian noise vector whose elements are i.i.d. with variance the data of sample k = 9 is added. The removed inducing point becomes
denoted by σx2 . In the proposed approach the squared expo- the RGP MS smoothed estimate.
nential covariance kernel [23] is adopted.
The measurements are stored and used for the GP
training. Training data older than the d most recent mea- B. Recursive Gaussian Process Motion Tracker (RGP MT)
surement samples, also called depth of the tracker, are The RGP MT is an online state estimation approach
assumed uncorrelated and removed. The position prediction for point target tracking. Consider again (42) and we will
and update equations for all output dimensions are similar. present the derivations for the x coordinate. Similar deriva-
The prediction and filtering of the x coordinate, using the GP tions hold true for the y coordinate and other states. The
regression (3) and (4) can be performed with the sequence GPMT assumptions 1, 2, 3, and 4 are also valid for the
of equations RGP MT. It is further assumed that the function values at
−1 the d most recent measurement samples are represented
μ̃xk = K t̄t k K̄ t k ztxk (44) by a set of d inducing points. This GP representation is
−1 different from the sparse GP of [29]. A bounded input
(φ̃kx )2 = K t̄ t̄ − K t̄t k K̄ t k K t̄tT k (45)
−1
domain is assumed in [29], whereas the input dimension
μ̂xk = K t̄t k K̄ t k ztx (46) for the proposed approach is right unbounded. Using 3, the
k
−1 input location of the inducing points are chosen to be same
(φ̂kx )2 = K t̄ t̄ − K kt k K̄ t k K t̄tT  (47)
k as that of the d most recent measurements and the input
locations of distant past are ignored.
where K̄a = [K aa + σx2 I d ], t̄ = k, t k = [k − d, k − d +
The target is assumed to follow an unknown nonlinear
1, . . . , k − 1]T , t k = [k − d + 1, k − d + 2, . . . , k ]T , μx
trajectory from time t = 0 to t = tmax . The input domain
and (φ x )2 denote, respectively, the position mean and vari-
is sampled using a fixed equidistant grid also called in-
ance. The GPMT prediction is explained in Fig. 1.
ducing points. Unlike [29] where multiple measurements
The derivative of a GP is also a GP [23]. The existence
are received at random input locations, the measurements
of a Derivative of a Gaussian Process (DGP) depends on
are received in a sequence starting at t = 0 and going
the differentiability of the covariance kernel. A squared
toward tmax . Moreover, a single measurement is received
exponential kernel is infinitely differentiable and as many
per sample. In order to keep the training computational
position derivatives can be predicted and filtered. Based on
complexity low and using three, a smaller set of inducing
the works of [33] and [34], a firstorder GPMT (FO-GPMT)
points (near the measured location) are considered for the
was proposed in [21] and is in the form
state prediction and filtering. In this way, the inducing points
−1 represent the nonlinear function in the near vicinity of the
μ̃ẋk = t̄ (K t̄t k )K̄ t k ztxk (48)
−1
measured location. This reduced set of inducing points is
(φ̃kẋ )2 = t̄2t̄ (K t̄ t̄ ) − t̄ (K t̄t k )K̄ t k [t̄ (K t̄t k ) ]T (49) then used for the functional prediction and filtering at the
−1 measured location. The processing time of the training was
μ̂ẋk = t̄ (K t̄t k )K̄ t  ztx (50)
k k reduced in the same way in GPMT [21]. However, the
−1
(φ̂kẋ )2 = t̄2t̄ (K t̄ t̄ ) − t̄ (K t̄t k )K̄ t  [t̄ (K t̄t k )]T (51) locations of the inducing points can be chosen different
k
from the measured locations. This is not possible in the
where μẋ and (φ ẋ )2 denote, respectively, the velocity mean GPMT [21]. Another advantage is that the location of the
and the variance. inducing points can be learned at each measurement sample

AFTAB AND MIHAYLOVA: LEARNING GAUSSIAN PROCESS APPROACH FOR MANEUVERING TARGET TRACKING 283
for optimal representation of the slice (of width d + 1) of of the inducing points, the prior at k is assumed to be
x xx
the underlying nonlinear function. As a result, the number p(xxk−1 |zx1:k−1 ) = N (μ̂xk−1 , Ĉ k−1 ) and that given the joint
of inducing points can be reduced when the target motion is distribution of the inducing points at t k and k, the likelihood
uniform. Conversely, when it is exhibiting sharp maneuvers, of the new measurement is independent of the previous
the inducing points can be increased. The whole process measurements. The posterior p(xxk |zx1:k ) can be updated
also employs online learning of the hyperparameters which using the decomposition [35] of the state into observable
makes the approach highly adaptive. x
oxk = [σkx , ( f̄ k )T ]T and unobservable uxk = [( f xk )T , (ηxk )T ]T
The inducing points are updated at each time step. parts
Consider the prior inducing points set locations at k are
represented by the d-dimensional prior input vector t k = p(xxk |zx1:k ) = p(uxk |oxk ) p(oxk |z1:k )d f̄kx . (55)
[k − d, k − d + 1, . . . , k − 1]T . The inducing points store
an estimate of the nonlinear function f x in the vector Inference is done using the RGP described in Section II-D.
f xk = f x (t k ). The prior inducing points are the training data The sigma points are chosen using the constrained un-
for the function prediction at t k and k. The measurement scented transform [36], which is different from [29]. The
zkx , received at k, updates the set of inducing points and the constraints are applied on the hyperparameters to remain
function prediction at the measured locations. The update positive at all times. The RGP recursion is adopted under
of the function at the measured location k is the position the following equivalent notations:
estimate of the target. The position estimate at k, evalu-
ated using the inducing point set and the measurement, is u = t k , ukf̄ = k, x̂k = xxk , f̄ k = f̄kx (56)
included in the inducing point set. The oldest, in terms of
zk = zkx , m(k ) = 0, m(t k ) = Od , J k = J k (57)
time, inducing point is removed and a new point is added to
the inducing point. As a result, the total number of inducing to give the RGP motion tracker (RGP MT). Using (56)
points is kept fixed. The new point is the function estimate at and (57) and the prior distribution of xx , the posterior is
k denoted as f̄kx = f x (k ). The recursion starts at k = d + 1 x
estimated using (21)–(41). The mean μ̂kf̄ and variance Ĉkf̄
x

sample. The initial distribution of the inducing points set and of the function at k can be found after modifying the update
the hyperparameters vector is determined using the first d (40) and (41) and the new expressions are
measurement samples, in the form
x
μ̂kf̄ = hT μ̂ok , h = 1Nk +1 − hT
T
x x (58)
p( f xd ) = N (μdf , C df = Ktd td + (σdx )2 I d ) (52) x o
Ĉkf̄ = h Ĉ k h
T
(59)
N (μθd , C θd )
x x
p(θ xd ) = (53)
x x
T o
x x C kf f̄ = h Ĉ k LTk (60)
where p(· ) is the probability density function, μdf and C df
x x
represent, respectively, the mean and xcovariance of the ini- where C kf f̄ represents the cross-covariance vector between
tial inducing point set and μθd and C θd denote, respectively,
x
the inducing points set and the estimate at the new input
the mean and covariance of the initial hyperparameters location and 1a is a 1 × a unit vector. Let the posterior mean
x
vector. The prior covariance C df differs from that proposed and covariance from (40) and (41) are
in [29] since the inducing points are initialized and re- x x f x f̄ x
cursively learned from noisy measurements. However, the μ̊kf = μkf , C̊kf = C kf , C̊ k f f̄
= Ck . (61)
inducing points are not learned exactly due to the measure-
ment noise and the uncertainty in the inducing points is Then, the posterior distribution of the inducing points can
represented with the following assumptions. be determined by this posterior mean and covariance given
in (61). Let the elements of the updated inducing point set,
1) The inducing points noise is modeled as an additive the covariance matrix x and the cross-covariance vector be
fx f f x f̄ x
random noise vector whose elements are i.i.d. with represented as μ̊k i , C̊ k i,i , and C̊ k i , respectively, where i =
variance (σdx )2 . {1, 2, . . . , d}. The most recent inducing point is denoted
2) The variance of the inducing point is of the order of with the index d. The following operations are performed
the measurement noise variance. on the updated inducing points set to get the modified set
As a result of the above assumptions, (10) is represented and to also keep the number of elements constant
as  fx
f xi μ̊k i+1 if i = 1, 2, . . . , d − 1
μ̂k = (62)
J k = K u f̄ u (K uu + (σkx )2 I d )−1 . (54) μ̂kf̄
x
ifi = d
k
⎧ f xi+1, j+1
The initial value of the mean vector μθd is found with the ⎪
x
⎪ C̊ if i, j ={1, . . . , d − 1}
⎪ kx x
⎪
ML estimation based hyperparameters optimization with ⎪
⎨ C̊ i+1 if i ={1, . . . , d − 1},j = d
f f̄
fx
the first d measurements. The inducing point vector and the Ĉ k i, j = k
(63)
hyperparameters vector are combined to define the state vec- ⎪
⎪ f xj+1 f̄ x
⎪
⎪ C̊ if i = d,j = { 1, . . . , d − 1}
tor xx = [( f x )T , (θ x )T ]T and θ x = [σ x , (ηx )T ]T . The state- ⎪ kx
⎩
space model is given by (21) and (22). For recursive update Ĉkf̄ if i = j = d.

284 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 1 FEBRUARY 2021
In (62) and (63), the oldest inducing point is removed and set and the derivative of the function at k are
a new inducing point is added, using the estimates and the x
ḟ
x
f
x
ḟ ˙f̄ x
corresponding cross-covariance matrix of the current time μ̊kḟ = μ̂kf , C̊ k = Ĉ k , C̊ k f f̄
= Ck
step. The state vector x defined in (21) and (22) is built using ˙x ˙x f̄
μ̂kf̄ = μ̂kf̄ , Ĉkf̄ = Ĉ k .
the new inducing point set and the updated hyperparameters
vector. The steps from (21) to (41) are repeated for the At the end of the recursion, the inducing points are updated,
estimates at the next sample. The RGP MT recursion is similarly to (62) and (63), as follows:
explained in Fig. 1. ⎧ x
x ⎨ μ̊ ḟ i+1 if i = 1, 2, . . . , d − 1
ḟ
C. Recursive Derivative of Gaussian Process Motion μ̂k i = k
(68)
⎩ μ̂ ˙f̄ x if i = d
Tracker (RDGP MT) k
⎧ ḟ x
The recursion for a DGP has not been proposed in ⎪
⎪ C̊ k
i+1, j +1
if i, j ={1, . . . , d − 1}
⎪
⎪
the literature. In this section, a recursive estimation of ⎪
⎪ x ˙x

the first-order derivative of the x position is proposed. In x ⎨ C̊ i+1 if i ={1, . . . , d − 1},j = d

ḟ f̄
ḟ
Ĉ k i, j = k
(69)
a similar way, the higher order derivatives of x and the ⎪
⎪
x
ḟ j +1 ˙f̄ x
⎪ C̊ k
⎪ if i = d,j = {1, . . . , d − 1}
time derivatives for other coordinates can be derived. This ⎪
⎪
formulation is generic and can be applied to any derivative ⎩ ˙f̄ x
Ĉk if i = j = d.
of a GP regression. Consider the GP model defined in (42).
The DGP is given as An alternative way of determining the derivatives (also used
in Section IV) is from the inducing points and the learned
dx df x (t ) hyperparameters of the RGP MT. The inducing points and
= = ḟ x (t ). (64)
dt dt derivatives can be represented by a multivariate Gaussian
A derivative of a GP is also a GP [23]. As described in distribution [34]
Section III-B, the input domain is sampled using a fixed     
f xk m(t k ) K t k t k + (σkx )2 I d t̄ (K t k t̄ )
˙f̄ x =N , .
grid of inducing points. The location of the inducing points
is chosen the same as for the RGP MT. Let the input vector k δ (m (t̄ )) t̄ (K t̄t k ) t̄2t̄ (Kt̄ t̄ )
of the inducing points is t k , then the local estimates of the (70)
DGP are stored in the following inducing points:
The predicted and filtered variables of the first-order deriva-
x x
ḟ = ḟ (t ). (65) tive process given the inducing points are
˙x −1 x
Consider the following initial distribution of the DGP: μ̃kf̄ = δ (m (t̄ )) + t̄ (K t̄t k )K̄ t k (μ̊kf − m(t k )) (71)
x x ˙x −1
x
pd ( ḟ ) = N (μdḟ , C dḟ = K t d t d + (σdẋ )2 I d ). (66) C̃kf̄ = t̄2t̄ (K t̄ t̄ ) − t̄ (K t̄t k )K̄ t k [t̄ (K t̄t k )]T (72)
˙f̄ x −1 fx
The posterior distribution is updated recursively using the μ̂k = δ (m (t̄ )) + t̄ (K t̄t k )K̄ t k (μ̂k − m(t k )) (73)
Bayes law and the Chapman–Kolmogrov equations given ˙x −1
Ĉkf̄ = t̄2t̄ (K t̄ t̄ ) − t̄ (K t̄t k )K̄ t k [t̄ (K t̄t k )]T . (74)
as
 x ¯x x ¯x x x x
x x p(żk | ḟ k , ḟ k ) p( ḟ k , ḟ k |ż1:k−1 )d ¯ḟ k D. Recursive Gaussian Process Motion
p( ḟ k |ż1:k ) = (67) Smoother (RGP MS) and Recursive Derivative of
ck
Gaussian Process Motion Smoother (RDGP MS)
x
where ż is the measurement, ¯ḟ k represents the function Training data for the GPMT prediction and update steps
evaluation at the measured location, and ck is the normal- are the recent measurements. Comparatively, the RGP MT
ization constant. It is assumed that the hyperparameters prediction is based on the inducing points set. The RGP MT
have been learned during the position filtering step, i.e., update step is performed using the inducing points set and
RGP MT. Hence, the posterior update is independent of the current measurement. The inducing points set is updated
the hyperparameters. The DGP inducing points and the during each measurement update. This update, given by
velocity inference and filtering at time k is done recursively (38) and (39), is similar to a fixed-lag-recursive-smoother.
following the RGP formulation given in Section II-B and The smoother lag is d measurement samples long according
the following equivalence: to (62) and (63). As shown in [29], this smoother is similar to
x x the augmented Kalman smoother of [37], with an additional
u = t k , ukf̄ = k, f k = ḟ k , f̄ k = ˙f̄ k , ṁ (k ) = 0 fx fx
measurement update step. μ̂k 1 and Ĉ k 1,1 in (62) and (63) are,
zx − zk−
x
2(σkx )2 respectively, the smoothed position and the corresponding
ṁ(t ) = Od , żkx = k 1
, (σ̇kx )2 = x x
T T variance evaluated at t  . Similarly, μ̂k 1 and Ĉ k 1,1 in (68)
ḟ ḟ

where T is the sampling time and żkx is a pseudomeasurement and (69) are, respectively, the smoothed velocity and the
with measurement noise variance (σ̇kx )2 . Using the RGP corresponding variance. As for the derivative GP, an alter-
recursion given in Section II-C, the filtered inducing points native way of smoothing is through the use of (73) and (74)

AFTAB AND MIHAYLOVA: LEARNING GAUSSIAN PROCESS APPROACH FOR MANEUVERING TARGET TRACKING 285
and setting t̄ = k − d and t k = [k, k − 1, . . . , k − d + 1]T . The process noise variance is set to qNCV =
The RGP MS is also explained in Fig. 1. 10 m2 /s4 .
F2) Fixed grid interacting multiple model (FGIMM).
E. Measurement Noise Uncertainty Analysis An FGIMM [43] is implemented using three KFs.
The proposed RGP MT and RGP MS does not require The grid consists of an NCV and two coordi-
prior knowledge of the measurement noise variances. This nated turn models. The rate of turns are set to
is achieved by setting the measurement variance as a hy- {−15, 15}◦ /s. The Markov transition probability of
perparameter. The measurement noise variance is recur- the same mode is set to 0.9 and for changing the
sively estimated at each step. As a result, the proposed mode is 0.05, the initial model probability vector is
filter and smoother are robust to the measurement noise {0.15, 0.7, 0.15} and the process noise variance is
errors, provided that the noise is additive Gaussian. The set to 26 m2 /s4 for each model. This is an optimum
performance of the proposed approaches is studied using process noise variance for a target moving with
different measurement noise variances and the results are 200 m/s according to [44].
presented in Section IV-F. F3) Singer. A KF designed using a Singer [20] state
transition model. The model parameters are set as
F. Sparsity and the Inducing Points follows; maximum possible acceleration is set to
Amax = 8 m/s2 , probability of no-acceleration is set
The RGP MT, RGP MS, RDGP MT, and RDGP MS to P0 = 0.4, probability of maximum acceleration
proposed in this article extend further the sparse GP [38], is set to Pmax = 0.1 and man oeuvre time constant
for target tracking and with studies on the impact of the is set to τm = 8 s.
noise parameters on the approaches’ performance. A typical F4) GPMT, FO-GPMT. A GPMT and FO-GPMT [21]
sparse GP application has a bounded input domain, whereas with depth set to d = 10 samples.
the proposed approaches in this article are based upon a F5) CGPMT, FOCGPMT. A constant (hyperparame-
right unbounded input domain. Second, a typical sparse ters) GPMT and FOGPMT with depth set to d =
GP application requires an estimate across the whole input 10 samples. The hyperparameters are initialized
domain. In target tracking, the interest lies at the current using the first d measurement samples and are kept
and future time, i.e., at a specific slice of the input domain. constant there after.
Finally, in a typical sparse GP application the training data F6) RGPMT, RDGPMT. A filter based on RGP. The
is available on both sides of the test data. Conversely, in learning is ML-based and the depth is set to d =
the assumed target filtering problem, the training data is not 10 samples.
available on the right side of the input domain. F7) RGP MT, RDGP MT, and RGP MS, RDGP MS.
The proposed approaches introduce sparsity through the The proposed filter and smoother with depth set to
inducing points and the parameter d. The inducing points d = 10 samples. The smoother is denoted as S*7.
locations are proposed to be the same as those of the mea-
surements. These can be chosen different from the measure- Filters (F1), (F2) and (F3) are chosen to compare the
ment locations as well. The distance between the inducing proposed approach against the model-based filtering meth-
points location cannot be kept too close or too far. The ods. Filter (F4) is chosen to study the effect of proposed
optimal grid points location is proposed as a function of the changes in the training (inducing points instead of mea-
lengthscale hyperparameter in [39]. Similarly, the accuracy surements) and learning (recursive instead of ML). Filter
of the developed approaches is sensitive to the parameter d. (F6), which adopts the same training process as the pro-
Previously, sparse GP methods have been proposed for the posed approach, is introduced to study the performance of
automatic selection of the inducing points [40] and choice different learning approaches, ML in (F6) versus recursive
of the parameter d [41]. A comprehensive review of these in the proposed (F10). Finally, filter (F5) is introduced to
methods is given in [42]. In this article, the parameter d is study the performance degradation if the hyperparameters
selected by the trial and error approach. In future, automatic are not learned.
selection can be introduced to improve the robustness of the
approach with respect to d. B. Testing Scenarios

IV. PERFORMANCE VALIDATION The above mentioned approaches are compared on the
following six challenging scenarios.
The proposed approach has been validated over chal-
lenging maneuvering scenarios presented in this section. S1) Uniform. The target velocity is kept nearly con-
stant. This scenario represents the commonly ob-
A. Compared Methods served trajectory of airliners.
Different model-based and model-free filters described S2) Gradual Coordinated Turns. The target motion
below are compared. is modeled using the left and right coordinated
turns (15◦ /s for 10 s) and the CV motion models.
F1) CV. A KF with state transition modeled us- This scenario represents maneuverable targets dy-
ing a nearly constant velocity (NCV) [2] model. namics, less agile than those in scenario (S3).

286 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 1 FEBRUARY 2021
 TABLE I
Match Mismatch Matrix

S3) Sharp Coordinated Turns. This is similar to (S2)

except the turn rates are set to (30◦ /s for 9 s). This
scenario represents highly maneuverable targets
dynamics such as military aircrafts.
S4) Singer Lazy. The target motion is modeled using
the Singer acceleration model. The parameters are
chosen the same as those of the Singer filter (F3).
S5) Singer Agile. This scenario is similar to (S4), except
the maximum acceleration is increased to A2max =
Fig. 2. Sample Trajectory. The figure shows a sample trajectory of each
50 m2 /s4 . As compared to (S4), an agile target is scenario. The initial position is indicated by a small circle. (a) S1: CV.
simulated in this test scenario. (b) S2: Gradual CT. (c) S3: Sharp CT. (d) S4: Singer lazy. (e) S5: Singer
S6) GP. The target motion is modeled using two zero Agile. (f) S6: GP.
mean GPs, each for the x and y coordinate, respec- 

tively. The squared exponential covariance kernel  1 
NMC 
K

is adopted for both GPs. The hyperparameters are mean rmse =  (qi,k − q̂i,k )2 (76)
KNMC i=1 k =1
kept constant. The magnitude of variance is set to
σgp
2
= 1e7 m2 and length-scale is set to lgp = 10 s.
where rmsekq and mean rmse represent, respectively, the
RMSE and the mean RMSE, qk is the true value, q̂k is the
The target velocity is initialized randomly in the
filter output, and K is the total number of samples.
range 150 ≤ v0 ≤ 250 m/s for model-based scenarios (S1)
to (S5). The total time of trajectory is 100 s. The mea- C. Implementation Details
surement noise standard deviation is set to σ = 25 m. The
simulated scenarios cover a wide range of maneuvering The implementation details of the proposed approach
trajectories depicted by the aerial targets. Each scenario is are given in this section. The state vector (excluding the hy-
matched, in terms of the structure and the parameters, to at- perparameters) and the measurement data are scaled down
least one of the filters. This provides a rigorous performance at the input and upscaled at the output of the filter. The
1
evaluation for each approach. The matrix for the match and scaling is set to 70 . The first two inducing points for the
mismatch among the filters and the scenarios is depicted in DGP are initialized to the same value. The mean hyper-
Table I. parameters vector is initialized using the ML of the first
Scenarios (S5) and (S6) represent highly maneuvering d measurement samples. Let the initial hyperparameters
target trajectories which are not matched to any of the vector is given as μθ0 = [σker 2
, l, σ 2 ]T , where σker
2
and l are
compared filters. A sample trajectory of each scenario is the kernel variance and lengthscale hyperparameters and
shown in Fig. 2. The filters are initialized using the mea- σ 2 is the noise variance hyperparameter. The correlation
surement data. The hyperparameters vector is initialized by between the noise variance hyperparameter and f (ukf̄ ) is
maximizing likelihood of first d measurement samples for necessary for its learning in (36) and (37). As proposed
filters (F6) and (F7). The position and velocity root mean in [29], this correlation is attained by correlating the noise
square errors (RMSE) are plotted in the figures showing variance hyperparameter with the remaining hyperparam-
performance graphs. The mean-RMSE errors of the position eters. To achieve this, the terms in the covariance of the
and velocity are given in the performance tables. The RMSE hyperparameters relating to the noise variance hyperparam-
and mean-RMSE in NMC Monte–Carlo simulation runs are eter matrix are set to nonzero values. The initial covariance
evaluated as matrix is ⎡ ⎤
0 σ40
2
 σker
2
 ⎢ 2⎥
 1  NMC
C θ0 = ⎣ 0 1 σ40 ⎦ . (77)
rmseq =
k  (qi,k − q̂i,k )2 (75) σ2 σ2 σ2
NMC i=1 40 40 20

AFTAB AND MIHAYLOVA: LEARNING GAUSSIAN PROCESS APPROACH FOR MANEUVERING TARGET TRACKING 287
 Fig. 3. Prediction RMSE. The figure shows the prediction performance in 10 000 Monte Carlo runs based on RMSE. An incomplete plot (such as
Fixed Grid Interacting Multiple Model (FGIMM) in S6) means that the corresponding filter sometimes diverges after that time. The Y-axis is set to log
scale for readability. (a) S1: CV. (b) S2: Gradual CT. (c) S3: Sharp CT. (d) S4: Singer lazy. (e) S5: Singer Agile. (f) S6: GP.

The cross-covariance matrix between the initial inducing TABLE II

points and the hyperparameters is set to a zero matrix Prediction Mean RMSE (NAN Means Filter Diverged)
fθ
Ĉ 0 = O.

D. Results
The results are obtained from 10 000 Monte Carlo runs
for each scenario. The accuracies of the prediction and
filtering steps are evaluated separately. The graphical and
numerical comparisons of the prediction process are given
in Fig. 3 and Table II, respectively. It can be observed that the
GPMT (F4), the proposed RGP MT (F7) and, a variation of
the proposed approach, RGPMT (F6) have comparable per-
formance in all scenarios. This shows that the performance (S5). This is due to the absence of learning in this filter. The
of real-time RGP MT, using the inducing points and online RMSE increases with time, which is not a desirable property
learning, is as good as ML-based nonreal-time GPMT. The for filters. The filter FGIMM (F2) sometimes diverges in
CGPMT (F5) performs poorly during the scenarios (S1) and scenario (S6), which is also not a desirable property. The

288 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 1 FEBRUARY 2021
Fig. 4. Filtering and Smoothing RMSE. The figure shows the filtering and smoothing performance in 10 000 Monte Carlo runs based on RMSE. An
incomplete plot (such as FGIMM in S6) means that the corresponding filter sometimes diverges. The Y-axis is set to log scale for readability. (a) S1:
CV. (b) S2: Gradual CT. (c) S3: Sharp CT. (d) S4: Singer lazy. (e) S5: Singer Agile. (f) S6: GP.

Singer KF (F3) performance degrades during the scenarios TABLE III

(S2), (S3), and (S6). The performance of the CV filter (F1) Filtering Mean RMSE (NAN Means Filter Diverged)
is not satisfactory for (S2), (S3), (S5), and (S6).
The graphical and numerical comparisons of the filter-
ing process are given in Fig. 4 and Table III, respectively. A
comparison similar to the prediction process can be done.
It can be concluded that the model-based filters and a
constant hyperparameters based GPMT are not suitable
for predicting a wide range of target trajectories. Three
highly maneuvering and mismatched scenarios, (S3), (S5),
and (S6), are considered to study the effects of the mismatch.
For scenario (S3), the proposed approach provides a per-
−109
formance improvement of 22109 × 100 = 80% in position
45−118 −37 −37
and 118 × 100 = 62% in velocity filtering as compared of 1937 × 100 = 49% in position and 2937 × 100 = 22%
to the best model-based filter (F2). For scenario (S5), the in velocity filtering as compared to the best model-based
proposed approach provides a performance improvement filter (F3).

AFTAB AND MIHAYLOVA: LEARNING GAUSSIAN PROCESS APPROACH FOR MANEUVERING TARGET TRACKING 289
 Fig. 5. Performance versus measurement noise variance. This figure shows the performance of the proposed approaches against different
measurement noise variances. The left and right halves correspond, respectively, to the prediction and the filtering processes. The numerics after the
name of the approach represents the scenarios, i.e., the postfix numeric 1 corresponds to 25 m standard deviation in measurement noise while 2 and 3
represent, respectively, 40 and 50 m standard deviations. (a) Prediction. (b) Filtering.

TABLE IV TABLE V
Processing Time in Milliseconds Percentage Degradation

is considered for the evaluation. The results are obtained

for three different measurement noise standard deviations,
For scenario (S6), the proposed approach provides a per-
−81 σx = {25, 40, 50} m. The position and velocity RMSE for
formance improvement of 2081 × 100 = 75% in position
27−64 5000 Monte Carlo runs of the three test scenarios are given
and 64 × 100 = 58% in velocity filtering as compared in Fig. 5. It can be observed that the performance degrades
to the best model-based filter (F3). The smoothing perfor- as the measurement noise variance increases. It is important
mance is also demonstrated in Fig. 4 and Table III. to note that no prior information regarding the change in the
noise variance is provided to the filter and smoother. The
E. Processing Time
proposed approaches adapt to the changing noise variance
The program was run on MATLAB R2018b on a Win- scenarios through recursive estimation of the variance and
dows 7 Home (64 bit) Laptop computer installed with an divergence is avoided. The percentage degradation in the
Intel(R) Core(TM) i3-M350 CPU @ 2.27GHz(2 CPUs) prediction, filtering, and the smoothing with respect to the
and 4GB RAM. The processing time of a single filter percentage increase in the noise variance is given in Table V.
iteration averaged over 500 Monte Carlo runs is given in It can be observed that the percentage degradation in the
Table IV, sorted from fastest to slowest. Filters (F4) and position prediction, filtering, and smoothing is comparable
(F6) take maximum time per iteration due to the time taken to the percentage increase in the noise variance, e.g., a 100%
for ML-based learning. These are not suitable for real-time increase in the noise variance degrades the position predic-
processing. Filters (F1), (F3), and (F5) are the fastest among tion by 92%. The percentage degradation in the velocity
the compared approaches. The combined processing time prediction, filtering, and smoothing is less as compared with
for the proposed filter and smoother is around 5ms. It is the percentage degradation of the position estimate, e.g., a
almost 10 times slower than (F2). The processing time 100% increase in the noise variance degrades the velocity
can be improved further by optimization of the code or by prediction by 42%.
changing the platform to C++.
V. CONCLUSION
F. Impact of Increased Noise Variance on the Model-based approaches are well established and
Performance widely used for state estimation. This article shows that
A simulation based study on the performance of the data-driven approaches can perform equally well and even
proposed approaches with an increasing noise variance is better than model-based approaches. A novel GP regres-
given in this section for x coordinate. Testing scenario (S6) sion approach with learning is proposed for tracking and

290 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 1 FEBRUARY 2021
smoothing purposes. Thanks to recursive learning of the [11] K. Granström, M. Baum, and S. Reuter
hyperparameters of the GP, the approach is suitable for Extended object tracking: Introduction, overview, and applica-
online estimation. The performance evaluation, carried out tions
J. Adv. Inf. Fusion, vol. 12, no. 2, pp. 139–174, 2017.
in challenging scenarios and compared with other online [12] D. Reid
model-based filters, shows that the proposed approach gives An algorithm for tracking multiple targets
overall the best performance for a wide set of target trajec- IEEE Trans. Autom. Control, vol. 24, no. 6, pp. 843–854, Dec.
tories. For the two highly maneuverable and mismatched 1979.
scenarios, (S3) CT and (S5) Singer, the proposed approach [13] S. Oh, S. Russell, and S. Sastry
Markov chain Monte Carlo data association for general
provides 80% and 62% performance improvement in the multiple-target tracking problems
position estimates and 49% and 22% in the velocity esti- in Proc. 43 rd IEEE Conf. Decis. Control, 2004, vol. 1, pp. 735–
mates, respectively, as compared to the best model-based 742.
filter (F3), i.e., a Singer motion model based KF. The [14] D. E. Clark
flexibility provided by the proposed model-free approach Multiple target tracking with the probability hypothesis density
filter
strongly advocates its preference over model-based ap- Ph.D. dissertation, Edinburgh, U.K.: Heriot-Watt Univ., 2006.
proaches for applications involving a wide set of target [15] R. E. Kalman
trajectories. Current work is focused on the performance A new approach to linear filtering and prediction problems
of GP approaches to quantify the impact of uncertainties J. Basic Eng., vol. 82, no. 1, pp. 35–45, 1960.
with information measures. [16] H. W. Sorenson
Kalman Filtering: Theory and Application. Piscataway, NJ,
USA: IEEE Press, 1985.
ACKNOWLEDGMENT
[17] E. A. Wan and R. Van Der Merwe
The authors would like to thank the Associate Editor and The unscented Kalman filter for nonlinear estimation
the anonymous reviewers who helped to improve this paper. in Proc. Adaptive Syst. Signal Process., Commun., Control
Symp., 2000, pp. 153–158.
[18] H. A. Blom and Y. Bar-Shalom
REFERENCES The interacting multiple model algorithm for systems with
Markovian switching coefficients
[1] Y. Bar-Shalom, T. E. Fortmann, and P. G. Cable
IEEE Trans. Autom. Control, vol. 33, no. 8, pp. 780–783, Aug.
Tracking and Data Association. New York, NY, USA: Elsevier,
1988.
1990.
[19] E. Mazor, A. Averbuch, Y. Bar-Shalom, and J. Dayan
[2] S. Blackman and R. Popoli
Interacting multiple model methods in target tracking: A survey
Design and Analysis of Modern Tracking Systems. Norwood,
IEEE Trans. Aerosp. Electron. Syst., vol. 34, no. 1, pp. 103–123,
MA, USA: Artech House, 1999.
Jan. 1998.
[3] N. N. Kachouie and P. W. Fieguth
[20] R. A. Singer
Extended-Hungarian-JPDA: Exact single-frame stem cell
Estimating optimal tracking filter performance for manned ma-
tracking
neuvering targets
IEEE Trans. Biomed. Eng., vol. 54, no. 11, pp. 2011–2019, Nov.
IEEE Trans. Aerosp. Electron. Syst., vol. AES-6, no. 4, pp. 473–
2007.
483, Jul. 1970.
[4] S. S. Blackman
[21] W. Aftab and L. Mihaylova
Multiple-Target Tracking With Radar Applications
A Gaussian process regression approach for point target track-
Dedham, MA, USA: Artech House, 1986, p. 463.
ing
[5] Y. Bar-Shalom, P. K. Willett, and X. Tian
in Proc. 22nd Int. Conf. Inf. Fusion, 2019, pp. 1–8.
Tracking and Data Fusion. Storrs, CT, USA: YBS Publishing,
[22] J. Hartikainen and S. Särkkä
2011.
Kalman filtering and smoothing solutions to temporal Gaussian
[6] Z.-M. Qian, X. E. Cheng, and Y. Q. Chen
process regression models
Automatically detect and track multiple fish swimming in shal-
in IEEE Int. Workshop Mach. Learn. Signal Process., 2010,
low water with frequent occlusion
pp. 379–384.
PLoS One, vol. 9, no. 9, 2014, Art. no. e106506.
[23] C. E. Rasmussen and C. K. Williams
[7] K. Granström, A. Natale, P. Braca, G. Ludeno, and F. Serafino
Gaussian Processes for Machine Learning, vol. 1, Cambridge,
Gamma Gaussian inverse Wishart probability hypothesis den-
MA, USA: MIT Press, 2006.
sity for extended target tracking using X-band marine radar data
[24] T. D. Barfoot, C. H. Tong, and S. Särkkä
IEEE Trans. Geosci. Remote Sens., vol. 53, no. 12, pp. 6617–
Batch continuous-time trajectory estimation as exactly sparse
6631, Dec. 2015.
Gaussian process regression
[8] A. Petrovskaya and S. Thrun
Robot., Sci. Syst., vol. 10, 2014, Paper A1.
Model based vehicle detection and tracking for autonomous
[25] J. Dong, B. Boots, and F. Dellaert
urban driving
Sparse Gaussian processes for continuous-time trajectory esti-
Auton. Robots, vol. 26, no. 2/3, pp. 123–139, 2009.
mation on matrix lie groups
[9] K. Wenzl, H. Ruser, and C. Kargel
2017, arXiv.1705.06020.
Configuration of a sparse network of LiDAR sensors to identify
[26] N. Wahlström and E. Özkan
security-relevant behavior of people
Extended target tracking using Gaussian processes
in Proc. SPIE Unmanned/Unattended Sensors Sensor Netw., E.
IEEE Trans. Signal Process., vol. 63, no. 16, pp. 4165–4178,
M. Carapezza, Ed., vol. 7480, 2009, pp. 48–57.
Aug. 2015.
[10] L. Mihaylova, A. Y. Carmi, F. Septier, A. Gning, S. K. Pang, and
[27] W. Aftab, A. De Freitas, M. Arvaneh, and L. Mihaylova
S. Godsill
A Gaussian process approach for extended object tracking with
Overview of Bayesian sequential Monte Carlo methods for
random shapes and for dealing with intractable likelihoods
group and extended object tracking
in Proc. 22nd Int. Conf. Digit. Signal Process., 2017, pp. 1–5.
Digit. Signal Process., vol. 25, pp. 1–16, 2014.

AFTAB AND MIHAYLOVA: LEARNING GAUSSIAN PROCESS APPROACH FOR MANEUVERING TARGET TRACKING 291
[28] M. Lázaro-Gredilla, S. Van Vaerenbergh, and N. D. Lawrence [37] S. Reece and S. Roberts
Overlapping mixtures of Gaussian processes for the data asso- An introduction to Gaussian processes for the Kalman filter
ciation problem expert
Pattern Recognit., vol. 45, no. 4, pp. 1386–1395, 2012. in Proc. 13th Int. Conf. Inf. Fusion, 2010, pp. 1–9.
[29] M. F. Huber [38] A. J. Smola and P. L. Bartlett
Recursive Gaussian process: On-line regression and learning Sparse greedy Gaussian process regression
Pattern Recognit. Lett., vol. 45, pp. 85–91, 2014. in Proc. Adv. Neural Inf. Process. Syst., 2001, pp. 619–625.
[30] S. J. Julier, J. K. Uhlmann, and H. F. Durrant-Whyte [39] C. Veibäck, J. Olofsson, T. R. Lauknes, and G. Hendeby
A new approach for filtering nonlinear systems Learning target dynamics while tracking using Gaussian pro-
in Proc. Amer. Control Conf., 1995, vol. 3, pp. 1628–1632. cesses
[31] X. R. Li and V. P. Jilkov IEEE Trans. Aerosp. Electron. Syst., vol. 56, no. 4,
Survey of maneuvering target tracking. Part I. Dynamic models pp. 2591–2602, Aug. 2020.
IEEE Trans. Aerosp. Electron. Syst., vol. 39, no. 4, pp. 1333– [40] E. Snelson and Z. Ghahramani
1364, Nov. 2003. Sparse Gaussian processes using pseudo-inputs
[32] O. Rudovic, M. Pantic, and I. Patras in Proc. Adv. Neural Inf. Process. Syst., 2006, pp. 1257–1264.
Coupled Gaussian processes for pose-invariant facial expres- [41] H.-M. Kim, B. K. Mallick, and C. Holmes
sion recognition Analyzing nonstationary spatial data using piecewise Gaussian
IEEE Trans. Pattern Anal. Mach. Intell., vol. 35, no. 6, pp. 1357– processes
1369, Jun. 2013. J. Amer. Statist. Assoc., vol. 100, no. 470, pp. 653–668, 2005.
[33] P. Boyle [42] H. Liu, Y. Ong, X. Shen, and J. Cai
Gaussian processes for regression and optimisation When Gaussian process meets big data: A review of scalable
Victoria Univ. Wellington, Wellington, New Zealand, 2007. GPs
[34] P. S. Swain et al. IEEE Trans. Neural Netw. Learn. Syst., pp. 1–19, 2020.
Inferring time derivatives including cell growth rates using [43] X. R. Li and V. P. Jilkov
Gaussian processes Survey of maneuvering target tracking. Part V. Multiple-model
Nature Commun., 2016, vol. 7, Art. no. 13766. methods
[35] F. Beutler, M. F. Huber, and U. D. Hanebeck IEEE Trans. Aerosp. Electron. Syst., vol. 41, no. 4, pp. 1255–
Gaussian filtering using state decomposition methods 1321, Oct. 2005.
in Proc. 12th Int. Conf. Inf. Fusion, 2009, pp. 579–586. [44] A. Munir and D. Atherton
[36] R. Kandepu, L. Imsland, and B. A. Foss Maneuvring target tracking using different turn rate models in
Constrained state estimation using the unscented Kalman filter the interacting multiple model algorithm
in Proc. 16th Mediterranean Conf. Control Autom., 2008, in Proc. 34th IEEE Conf. Decis. Control, 1995, vol. 3, pp. 2747–
pp. 1453–1458. 2751.

Waqas Aftab received the B.Eng. degree in avionics from the College of Aeronautical
Engineering (NUST), Risalpur, Pakistan, in 2005, and the M.S. degree in control systems
engineering from Air University, Islamabad, Pakistan, in 2014.
He is a Ph.D. Research Student with The University of Sheffield, Sheffield, U.K. His
main interests are multisensor multitarget tracking and Gaussian processes.

Lyudmila Mihaylova (Senior Member, IEEE) received the M.Eng. degree in systems and
control engineering, the M.Sc. degree in applied mathematics and informatics, and the
Ph.D. degree in systems and control engineering from the Technical University of Sofia,
Bulgaria.
She is currently a Professor of Signal Processing and Control with the Department
of Automatic Control and Systems Engineering, University of Sheffield, Sheffield, U.K.
Her research interests include machine learning and autonomous systems with various
applications such as navigation, surveillance, and sensor network systems. She is an
Associate Editor for the IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS
and for the Elsevier Signal Processing Journal. She was the President of the International
Society of Information Fusion (ISIF) in the period 2016–2018. She is on the Board of
Directors of ISIF. She was the General Vice-Chair for the International Conference on
Information Fusion 2018 (Cambridge, U.K.), of the IET Data Fusion and Target Tracking
2014 and 2012 Conferences, Program Co-Chair for the 19th International Conference on
Information Fusion 2020, 2016, and others.

292 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 1 FEBRUARY 2021