Improved Adaptive Kalman Filter With Unknown
Process Noise Covariance
Jirong Ma
1,2
, Hua Lan
1,2
, Zengfu Wang
1,2
, Xuezhi Wang
3
, Quan Pan
1,2
, and Bill Moran
4
1
School of Automation, Northwestern Polytechnical University, China
2
Key laboratory of Information Fusion Technology, Ministry of Education, China
Email: mjr@mail.nwpu.edu.cn, {lanhua, wangzengfu, quanpan}@nwpu.edu.cn
3
School of Engineering, RMIT University, Australia
Email: xuezhi.wang@rmit.edu.au
4
Dept. Electrical and Electronic Engineering, University of Melbourne, Australia
Email: wmoran@unimelb.edu.au
Abstract—This paper considers the joint recursive estimation
of the dynamic state and the time-varying process noise covari-
ance for a linear state space model. The conjugate prior on
the process noise covariance, the inverse Wishart distribution,
provides a latent variable. A variational Bayesian inference
framework is then adopted to iteratively estimate the posterior
density functions of the dynamic state, process noise covariance
and the introduced latent variable. The performance of the
algorithm is demonstrated with simulated data in a target
tracking application.
Index Terms—Adaptive filtering, unknown process noise co-
variance, variational Bayesian, conjugate priors
I. INTRODUCTION
Linear state-space models are almost ubiquitously used
to represent real world dynamical systems, where possible
system disturbance or time-evolution variation is modeled
by system process noise. When the Kalman filter (KF) is
exploited to estimate the state of such systems, a lack of
knowledge of process noise statistics is known to result in
estimation error [1]. Such a situation occurs, for example,
in autonomous navigation and target tracking, where large
uncertainty of system evolution may appear and the process
noise covariance varies with time. Adaptive filtering, which
estimates both the unknown model parameters and system
dynamic state simultaneously from measurements, is an ideal
approach to address this issue [2]. State-of-the-art adaptive
filtering approaches may be categorized into four methods,
Bayesian, maximum likelihood, correlation, and covariance
matching [1]. State augmentation [3], interactive multiple
models [4] and particle filters [5] are the most well-known
Bayesian approaches. The safe-husa adaptive KF of [6] is a
covariance matching method, which estimates process noise
covariance and measurement noise covariance sequentially.
Variational Bayesian inference (VB) is a closed-loop it-
erative method that is often able to turn difficult inference
problems into optimization problems, and has lower computa-
tional cost compared to, for instance, a randomized sampling
method [7], [8]. In recent years, VB has been applied to
adaptive filtering through joint state and noise covariance
estimation. For the linear state space model, Sakka and Num-
menmaa [2] considered an unknown diagonal measurement
noise covariance matrix, each element of which is assumed
to follow an inverse Gamma distribution, as the conjugate
prior for the variance of a Gaussian distribution. Then, an
adaptive KF method based on VB (VBAKF) was used for
the joint estimation of the dynamic state and the measurement
noise covariance. This work was further extended in [9], where
an inverse Wishart prior is used for a general (not diagonal)
measurement covariance matrix with nonlinear system dynam-
ics, and in [10], where a t-distribution rather than a Gaussian
is used for the measurement distribution to improve outlier
elimination. The latter paper also uses a Rauch-Tung-Striebel
smoother for state estimation. However, extending the work
of [9] to consider unknown process noise covariance is not
easy since the process noise covariance does not appear in as
straightforward conjugate prior form as the measurement noise
[9].
Ardeshiri et al. [11] presented a batch-processing VB al-
gorithm for joint estimation of the dynamic system state,
the measurement noise covariance and the process noise
covariance, with the noise covariance matrices being identified
off-line. Instead of estimating the process noise covariance
matrix directly, Huang et al. [12] proposed a novel VBAKF
in which the inverse Wishart distribution was used as a prior
for the predicted error covariance matrix and measurement
noise covariance matrix, and inferred the system state with
the unknown covariance matrices of process noise and mea-
surement noise. The approach of [12] is an on-line recursive
method but it needs a nominal process noise covariance matrix
at each time step as the parameter of the algorithm.
In this paper, we develop a novel VB based adaptive
KF algorithm, referred as VBAKF-Q, for state estimation
with unknown process noise covariance. By introducing a
new latent variable, the conjugate prior distribution of the
underlying process noise covariance is assumed to follow
the inverse Wishart distribution. Thus, the problem of joint