Professional Documents
Culture Documents
Prediction step
Prior knowledge
of state Based on e.g. can run in real time, using only the present input measure-
physical model
ments and the previously calculated state and its uncer-
Next timestep tainty matrix; no additional past information is required.
The Kalman lter does not make any assumption that the
Update step
Compare prediction
Measurements
errors are Gaussian.[3] However, the lter yields the exact
to measurements conditional probability estimate in the special case that all
Output estimate errors are Gaussian-distributed.
of state
1
2 4 TECHNICAL DESCRIPTION AND CONTEXT
KalmanBucy lter because it is a special case of a more volved in a single set of calculations. This allows for a rep-
general, non-linear lter developed somewhat earlier by resentation of linear relationships between dierent state
the Soviet mathematician Ruslan Stratonovich.[8][9][10][11] variables (such as position, velocity, and acceleration) in
In fact, some of the special case linear lters equa- any of the transition models or covariances.
tions appeared in these papers by Stratonovich that were
published before summer 1960, when Kalman met with
Stratonovich during a conference in Moscow.
3 Example application
As an example application, consider the problem of de-
2 Overview of the calculation termining the precise location of a truck. The truck can
be equipped with a GPS unit that provides an estimate of
the position within a few meters. The GPS estimate is
The Kalman lter uses a systems dynamics model (e.g.,
likely to be noisy; readings 'jump around' rapidly, though
physical laws of motion), known control inputs to that
remaining within a few meters of the real position. In
system, and multiple sequential measurements (such as
addition, since the truck is expected to follow the laws
from sensors) to form an estimate of the systems varying
of physics, its position can also be estimated by integrat-
quantities (its state) that is better than the estimate ob-
ing its velocity over time, determined by keeping track
tained by using only one measurement alone. As such, it
of wheel revolutions and the angle of the steering wheel.
is a common sensor fusion and data fusion algorithm.
This is a technique known as dead reckoning. Typically,
Noisy sensor data, approximations in the equations that the dead reckoning will provide a very smooth estimate
describe the system evolution, and external factors that of the trucks position, but it will drift over time as small
are not accounted for all place limits on how well it is pos- errors accumulate.
sible to determine the systems state. The Kalman lter
In this example, the Kalman lter can be thought of as op-
deals eectively with the uncertainty due to noisy sensor
erating in two distinct phases: predict and update. In the
data and to some extent also with random external factors.
prediction phase, the trucks old position will be modied
The Kalman lter produces an estimate of the state of the
according to the physical laws of motion (the dynamic or
system as an average of the systems predicted state and of
state transition model). Not only will a new position
the new measurement using a weighted average. The pur-
estimate be calculated, but a new covariance will be cal-
pose of the weights is that values with better (i.e., smaller)
culated as well. Perhaps the covariance is proportional
estimated uncertainty are trusted more. The weights
to the speed of the truck because we are more uncer-
are calculated from the covariance, a measure of the esti-
tain about the accuracy of the dead reckoning position
mated uncertainty of the prediction of the systems state.
estimate at high speeds but very certain about the posi-
The result of the weighted average is a new state estimate
tion estimate when moving slowly. Next, in the update
that lies between the predicted and measured state, and
phase, a measurement of the trucks position is taken from
has a better estimated uncertainty than either alone. This
the GPS unit. Along with this measurement comes some
process is repeated at every time step, with the new esti-
amount of uncertainty, and its covariance relative to that
mate and its covariance informing the prediction used in
of the prediction from the previous phase determines how
the following iteration. This means that the Kalman lter
much the new measurement will aect the updated pre-
works recursively and requires only the last best guess,
diction. Ideally, as the dead reckoning estimates tend to
rather than the entire history, of a systems state to calcu-
drift away from the real position, the GPS measurement
late a new state.
should pull the position estimate back towards the real po-
The relative certainty of the measurements and current sition but not disturb it to the point of becoming rapidly
state estimate is an important consideration, and it is com- jumping and noisy.
mon to discuss the response of the lter in terms of the
Kalman lters gain. The Kalman gain is the relative
weight given to the measurements and current state es-
timate, and can be tuned to achieve particular perfor- 4 Technical description and con-
mance. With a high gain, the lter places more weight text
on the most recent measurements, and thus follows them
more responsively. With a low gain, the lter follows the The Kalman lter is an ecient recursive lter that
model predictions more closely. At the extremes, a high estimates the internal state of a linear dynamic system
gain close to one will result in a more jumpy estimated from a series of noisy measurements. It is used in a
trajectory, while low gain close to zero will smooth out wide range of engineering and econometric applications
noise but decrease the responsiveness. from radar and computer vision to estimation of struc-
When performing the actual calculations for the lter (as tural macroeconomic models,[12][13] and is an important
discussed below), the state estimate and covariances are topic in control theory and control systems engineering.
coded into matrices to handle the multiple dimensions in- Together with the linear-quadratic regulator (LQR), the
3
by user 0, Q 0, Q 0, Q
the entire internal state. F B H F B H F B H
wk -1 vk -1 wk vk wk +1 vk+1
Hidden
In DempsterShafer theory, each state equation or obser- xk-1,Pk -1 xk ,Pk xk+1,Pk +1
ALS, mentioned in the section above. After the covari- an unknown input and G applies that eect to the state
ances are estimated, it is useful to evaluate the perfor- vector) where
mance of the lter; i.e., whether it is possible to improve
the state estimation quality. If the Kalman lter works [ ]
optimally, the innovation sequence (the output prediction 1 t
F=
error) is a white noise, therefore the whiteness property 0 1
of the innovations measures lter performance. Several [ 1 2]
t
dierent methods can be used for this purpose.[23] If the G = 2
noise terms are non-Gaussian distributed, methods for t
assessing performance of the lter estimate, which use
probability inequalities or large-sample theory, are given so that
in [24] and.[25]
xk = Fxk1 + wk
7 Example application, technical where
wk N (0, Q)
[1 4 1 3
]
4 t 2 t
Q= GGT a2 = a2 .
1 3 2
2 t t
Please note that the matrix Q is not full rank (it is of rank
one if t = 0 ). Hence, the distribution N (0, Q) is
not absolutely continuous and has no probability density
function. Another way to express this, avoiding explicit
degenerate distributions is given by
Truth; ltered process; observations.
(note that there is no Bu term since we have no known and to tell the lter that we know the exact position and
control inputs. Instead, we assume that ak is the eect of velocity, we give it a zero covariance matrix:
6 8 DERIVATIONS
The lter will then prefer the information from the rst 8.2 Kalman gain derivation
measurements over the information already in the model.
The Kalman lter is a minimum mean-square error esti-
mator. The error in the a posteriori state estimation is
8 Derivations
x x
8.1 Deriving the a posteriori estimate co- k k|k
variance matrix We seek to minimize the expected
[ value of the square
] of
2
the magnitude of this vector, E
xk x k|k
. This
Starting with our invariant on the error covariance Pk | k
as above is equivalent to minimizing the trace of the a posteriori
estimate covariance matrix Pk|k . By expanding out the
terms in the equation above and collecting, we get:
( )
Pk|k = cov xk x
k|k
and by collecting the error vectors we get Kk Sk = (Hk Pk|k1 )T = Pk|k1 HTk
Kk = Pk|k1 HTk S1
k
[( ) ]
Pk|k = cov I Kk Hk )(xk x
k|k1 Kk vk This gain, which is known as the optimal Kalman gain, is
the one that yields MMSE estimates when used.
Since the measurement error vk is uncorrelated with the
other terms, this becomes
8.3 Simplication of the a posteriori error
[ ( )] covariance formula
Pk|k = cov (I Kk Hk ) xk x
k|k1 + cov [Kk vk ]
The formula used to calculate the a posteriori error co-
by the properties of vector covariance this becomes variance can be simplied when the Kalman gain equals
the optimal value derived above. Multiplying both sides
( ) of our Kalman gain formula on the right by SkKkT , it
T
Pk|k = I Kk Hk ) cov(xk x
k|k1 (I Kk Hk ) +Kk follows KTk
cov (vk )that
7
Pk|k = Pk|k1 Kk Hk Pk|k1 Pk|k1 HTk KTk +Kk Sk KTk Pak|k1 = Fk Pak1|k1 FTk + Qak
nonsingular matrix.[31] Any singular covariance matrix is p(x0 , . . . , xk , z1 , . . . , zk ) = p(x0 ) p(zi | xi )p(xi | xi1 )
i=1
pivoted so that the rst diagonal partition is nonsingular
and well-conditioned. The pivoting algorithm must retain However, when the Kalman lter is used to estimate the
any portion of the innovation covariance matrix directly state x, the probability distribution of interest is that as-
corresponding to observed state-variables H x | - that sociated with the current states conditioned on the mea-
are associated with auxiliary observations in y . The ldlt surements up to the current timestep. This is achieved by
square-root lter requires orthogonalization of the obser- marginalizing out the previous states and dividing by the
vation vector.[29][30] This may be done with the inverse probability of the measurement set.
square-root of the covariance matrix for the auxiliary
variables using Method 2 in Higham (2002, p. 263).[32] This leads to the predict and update steps of the Kalman
lter written probabilistically. The probability distribu-
tion associated with the predicted state is the sum (inte-
11 Relationship to recursive gral) of the products of the probability distribution asso-
ciated with the transition from the (k 1)-th timestep to
Bayesian estimation the k-th and the probability distribution associated with
the previous state, over all possible xk1 .
The Kalman lter can be presented as one of the sim-
plest dynamic Bayesian networks. The Kalman lter cal-
culates estimates of the true values of states recursively
p(xk | Zk1 ) = p(xk | xk1 )p(xk1 | Zk1 ) dxk1
over time using incoming measurements and a mathemat-
ical process model. Similarly, recursive Bayesian estima-
tion calculates estimates of an unknown probability den- The measurement set up to time t is
sity function (PDF) recursively over time using incoming
measurements and a mathematical process model.[33]
In recursive Bayesian estimation, the true state is assumed Zt = {z1 , . . . , zt }
to be an unobserved Markov process, and the measure-
ments are the observed states of a hidden Markov model The probability distribution of the update is proportional
(HMM). to the product of the measurement likelihood and the pre-
dicted state.
The denominator
hidden markov model
p(zk | Zk1 ) = p(zk | xk )p(xk | Zk1 ) dxk
because of the Markov assumption, the true state is con-
ditionally independent of all earlier states given the im- is a normalization term.
mediately previous state.
The remaining probability density functions are
12 Marginal likelihood
T
Related to the recursive Bayesian interpretation described p(z) = p(zk | xk )p(xk | zk1 , . . . , z0 )dxk
above, the Kalman lter can be viewed as a generative k=0
T
model, i.e., a process for generating a stream of random
observations z = (z0 , z1 , z2 , ). Specically, the process = N (zk ; Hk xk , Rk )N (xk ; x
k|k1 , Pk|k1 )dxk
k=0
is
T
= N (zk ; Hk x
k|k1 , Rk + Hk Pk|k1 HTk )
k=0
1. Sample a hidden state x0 from the Gaussian prior
distribution p(x0 ) = N (
x0|0 , P0|0 ) .
T
= N (zk ; Hk x
k|k1 , Sk ),
k=0
2. Sample an observation z0 from the observation i.e., a product of Gaussian densities, each correspond-
model p(z0 | x0 ) = N (H0 x0 , R0 ) . ing to the density of one observation zk under the cur-
rent ltering distribution Hk x
k|k1 , Sk . This can easily
3. For k = 1, 2, 3, . . . , do be computed as a simple recursive update; however, to
avoid numeric underow, in a practical implementation
it is usually desirable to compute the log marginal like-
(a) Sample the next hidden state xk from the tran- lihood = log p(z) instead. Adopting the convention
sition model p(xk | xk1 ) = N (Fk xk1 + (1) = 0 , this can be done via the recursive update rule
Bk uk , Qk ).
(b) Sample an observation zk from the observation (k) = (k1) 1 (yT S1 y + log |S | + d log 2 ) ,
k k y
model p(zk | xk ) = N (Hk xk , Rk ). 2 k k
where dy is the dimension of the measurement vector.[34]
Note that this process has identical structure to the hidden An important application where such a (log) likelihood
Markov model, except that the discrete state and obser- of the observations (given the lter parameters) is used
vations are replaced with continuous variables sampled is multi-target tracking. For example, consider an ob-
from Gaussian distributions. ject tracking scenario where a stream of observations is
the input, however, it is unknown how many objects are
In some applications, it is useful to compute the prob- in the scene (or, the number of objects is known but is
ability that a Kalman lter with a given set of parame- greater than one). In such a scenario, it can be unknown
ters (prior distribution, transition and observation mod- apriori which observations/measurements were generated
els, and control inputs) would generate a particular ob- by which object. A multiple hypothesis tracker (MHT)
served signal. This probability is known as the marginal typically will form dierent track association hypothe-
likelihood because it integrates over (marginalizes out) ses, where each hypothesis can be viewed as a Kalman
the values of the hidden state variables, so it can be lter (in the linear Gaussian case) with a specic set of pa-
computed using only the observed signal. The marginal rameters associated with the hypothesized object. Thus,
likelihood can be useful to evaluate dierent parameter it is important to compute the likelihood of the obser-
choices, or to compare the Kalman lter against other vations for the dierent hypotheses under consideration,
models using Bayesian model comparison. such that the most-likely one can be found.
It is straightforward to compute the marginal likelihood
as a side eect of the recursive ltering computation. By
the chain rule, the likelihood can be factored as the prod- 13 Information lter
uct of the probability of each observation given previous
observations, In the information lter, or inverse covariance lter, the
estimated covariance and estimated state are replaced by
the information matrix and information vector respec-
T tively. These are dened as:
p(z) = p(zk | zk1 , . . . , z0 )
k=0
Yk|k = P1
k|k
and because the Kalman lter describes a Markov pro- y = P1 x
k|k k|k k|k
cess, all relevant information from previous observations
k|k1 , Pk|k1 . Similarly the predicted covariance and state have equiva-
is contained in the current state estimate x
Thus the marginal likelihood is given by lent information forms, dened as:
10 15 FIXED-INTERVAL SMOOTHERS
where:
Yk|k1 = P1
k|k1
x
t|t1 is estimated via a standard Kalman lter;
yk|k1 = P1
k|k1 xk|k1
yt|t1 = zt H xt|t1 is the innovation produced
as have the measurement covariance and measurement
considering the estimate of the standard Kalman l-
vector, which are dened as:
ter;
the various xti|t with i = 1, . . . , N 1 are new
Ik = HTk R1
k Hk variables; i.e., they do not appear in the standard
ik = HTk R1
k zk Kalman lter;
The information update now becomes a trivial sum.[35] the gains are computed via the following scheme:
Yk|k = Yk|k1 + Ik
[ ]1
yk|k = yk|k1 + ik K(i) = P(i) HT HPHT + R
The main advantage of the information lter is that N and
measurements can be ltered at each timestep simply by [ ]i
summing their information matrices and vectors. P(i) = P (F KH)
T
Yk|k1 = Lk Mk LTk + Ck Q1 T
k Ck
[ 1 ]T 15 Fixed-interval smoothers
yk|k1 = Lk Fk yk1|k1
Note that if F and Q are time invariant these values can
The optimal xed-interval smoother provides the optimal
be cached. Note also that F and Q need to be invertible.
k|n ( k < n ) using the measurements from
estimate of x
a xed interval z1 to zn . This is also called Kalman
Smoothing. There are several smoothing algorithms in
14 Fixed-lag smoother common use.
time step and proceed backwards in time using the fol- 15.3 Minimum-variance smoother
lowing recursive equations:
The minimum-variance smoother can attain the best-
possible error performance, provided that the models are
x
k|n = x xk+1|n x
k|k + Ck ( k+1|k ) linear, their parameters and the noise statistics are known
precisely.[38] This smoother is a time-varying state-space
P = P + C (P P )CT generalization of the optimal non-causal Wiener lter.
k|n k|k k k+1|n k+1|k k
The smoother calculations are done in two passes. The
where forward calculations involve a one-step-ahead predictor
and are given by
Ck = Pk|k FTk+1 P1
k+1|k
16 Frequency-weighted Kalman the Kalman lter equations. This process essentially lin-
earizes the non-linear function around the current esti-
lters mate.
The basic Kalman lter is limited to a linear assumption. As with the EKF, the UKF prediction can be used inde-
More complex systems, however, can be nonlinear. The pendently from the UKF update, in combination with a
non-linearity can be associated either with the process linear (or indeed EKF) update, or vice versa.
model or with the observation model or with both.
The estimated state and covariance are augmented with
the mean and covariance of the process noise.
17.1 Extended Kalman lter
[ [ ]]T
Main article: Extended Kalman lter xak1|k1 = xTk1|k1 E wTk
[ ]
a Pk1|k1 0
In the extended Kalman lter (EKF), the state transition Pk1|k1 =
0 Qk
and observation models need not be linear functions of
the state but may instead be non-linear functions. These A set of 2L + 1 sigma points is derived from the aug-
functions are of dierentiable type. mented state and covariance where L is the dimension of
the augmented state.
xk = f (xk1 , uk ) + wk
zk = h(xk ) + vk 0k1|k1 = xak1|k1
( )
The function f can be used to compute the predicted state ik1|k1 = xak1|k1 + (L + )Pak1|k1 , i = 1, . . . , L
from the previous estimate and similarly the function h ( )i
can be used to compute the predicted measurement from k1|k1 = xk1|k1
i a
(L + )Pak1|k1 , i = L + 1, . . .
iL
the predicted state. However, f and h cannot be applied to
the covariance directly. Instead a matrix of partial deriva- where
tives (the Jacobian) is computed.
At each timestep the Jacobian is evaluated with cur- ( )
rent predicted states. These matrices can be used in (L + )Pak1|k1
i
17.2 Unscented Kalman lter 13
is the ith column of the matrix square root of As before, a set of 2L + 1 sigma points is derived from the
augmented state and covariance where L is the dimension
of the augmented state.
(L + )Pak1|k1
2L The sigma points are projected through the observation
x
k|k1 = Wsi ik|k1 function h.
i=0
2L [ ][ ]T
Pk|k1 = Wci ik|k1 x
k|k1 ik|k1 x
k|k1 i = h(i
k k|k1 ) i = 0, . . . , 2L
i=0
The weighted sigma points are recombined to produce
where the weights for the state and covariance are given the predicted measurement and predicted measurement
by: covariance.
2L
Ws0 =
L+ zk = Wsi ki
i=0
0
Wc = + (1 2 + )
L+
2L
[ ][ ]T
1 Pzk zk = Wci ki zk ki zk
Wsi = Wci = i=0
2(L + )
= 2 (L + ) L The state-measurement cross-covariance matrix,
where
Pk|k = Pk|k1 Kk Pzk zk KkT
xk = x(tk )
18 KalmanBucy lter
Initialize x
0|0 = E [x(t0 )] , P0|0 = Var [x(t0 )]
The KalmanBucy lter (named after Richard Snow-
den Bucy) is a continuous time version of the Kalman
lter.[46][47] (t) = F(t)
x x(t) + B(t)u(t) with ,
x(tk1 ) = x
k1|k1
It is based on the state space model x
k|k1 = x (tk )
Predict
P(t) = F(t)P(t) + P(t)F(t)T + Q(t) with ,P(tk1 ) = Pk
Pk|k1 = P(tk )
d
x(t) = F(t)x(t) + B(t)u(t) + w(t)
dt
The prediction equations are derived from those of
z(t) = H(t)x(t) + v(t)
continuous-time Kalman lter without update from mea-
where Q(t) and R(t) represent the intensities (or, more surements, i.e., K(t) = 0 . The predicted state and co-
accurately: the Power Spectral Density - PSD - matrices) variance are calculated respectively by solving a set of
of the two white noise terms w(t) and v(t) , respectively. dierential equations with the initial value equal to the
estimate at the previous step.
The lter consists of two dierential equations, one for
the state estimate and one for the covariance: ( )1
Kk = Pk|k1 HTk Hk Pk|k1 HTk + Rk
Update x k|k1 + Kk (zk Hk x
k|k = x k|k1 )
d Pk|k = (I Kk Hk )Pk|k1
x
(t) = F(t)x(t) + B(t)u(t) + K(t) (z(t) H(t) x(t))
dt
d
P(t) = F(t)P(t) + P(t)FT (t) + Q(t) K(t)R(t)KT (t) The update equations are identical to those of the
dt discrete-time Kalman lter.
where the Kalman gain is given by
Invariant extended Kalman lter [8] Stratonovich, R. L. (1959). Optimum nonlinear systems
which bring about a separation of a signal with constant
Kernel adaptive lter parameters from noise. Radiozika, 2:6, pp. 892901.
16 23 REFERENCES
[9] Stratonovich, R. L. (1959). On the theory of optimal non- [24] Spall, James C. (1995). The Kantorovich inequal-
linear ltering of random functions. Theory of Probability ity for error analysis of the Kalman lter with un-
and its Applications, 4, pp. 223225. known noise distributions. Automatica. 31 (10): 1513.
doi:10.1016/0005-1098(95)00069-9.
[10] Stratonovich, R. L. (1960) Application of the Markov pro-
cesses theory to optimal ltering. Radio Engineering and [25] Maryak, J.L.; Spall, J.C.; Heydon, B.D. (2004).
Electronic Physics, 5:11, pp. 119. Use of the Kalman Filter for Inference in State-
Space Models with Unknown Noise Distributions.
[11] Stratonovich, R. L. (1960). Conditional Markov Pro- IEEE Transactions on Automatic Control. 49: 87.
cesses. Theory of Probability and its Applications, 5, pp. doi:10.1109/TAC.2003.821415.
156178.
[26] Anderson, Brian D. O.; Moore, John B. (1979). Optimal
[12] Ingvar Strid; Karl Walentin (April 2009). Block Filtering. New York: Prentice Hall. pp. 129133. ISBN
Kalman Filtering for Large-Scale DSGE Models. Com- 0-13-638122-7.
putational Economics. Springer. 33 (3): 277304.
doi:10.1007/s10614-008-9160-4. [27] Jingyang Lu. False information injection attack on dy-
namic state estimation in multi-sensor systems, Fusion
[13] Martin Mller Andreasen (2008). Non-linear DSGE 2014
Models, The Central Dierence Kalman Filter, and The
Mean Shifted Particle Filter (PDF). [28] Thornton, Catherine L. (15 October 1976). Triangular
Covariance Factorizations for Kalman Filtering (PDF).
[14] Roweis, S; Ghahramani, Z (1999). A unifying review (PhD thesis). NASA. NASA Technical Memorandum 33-
of linear gaussian models. Neural computation. 11 798.
(2): 30545. doi:10.1162/089976699300016674. PMID
9950734. [29] Bierman, G.J. (1977). Factorization Methods for Dis-
crete Sequential Estimation. Factorization Methods
[15] Hamilton, J. (1994), Time Series Analysis, Princeton Uni- for Discrete Sequential Estimation. Academic Press.
versity Press. Chapter 13, 'The Kalman Filter' Bibcode:1977fmds.book.....B.
[16] Ishihara, J.Y.; Terra, M.H.; Campos, J.C.T. (2006). [30] Bar-Shalom, Yaakov; Li, X. Rong; Kirubarajan, Thia-
Robust Kalman Filter for Descriptor Systems. IEEE galingam (July 2001). Estimation with Applications to
Transactions on Automatic Control. 51 (8): 1354. Tracking and Navigation. New York: John Wiley & Sons.
doi:10.1109/TAC.2006.878741. pp. 308317. ISBN 978-0-471-41655-5.
[17] Terra, Marco H.; Cerri, Joao P.; Ishihara, Joao Y. [31] Golub, Gene H.; Van Loan, Charles F. (1996). Matrix
(2014). Optimal Robust Linear Quadratic Regulator Computations. Johns Hopkins Studies in the Mathemat-
for Systems Subject to Uncertainties. IEEE Trans- ical Sciences (Third ed.). Baltimore, Maryland: Johns
actions on Automatic Control. 59 (9): 25862591. Hopkins University. p. 139. ISBN 978-0-8018-5414-9.
doi:10.1109/TAC.2014.2309282.
[32] Higham, Nicholas J. (2002). Accuracy and Stability of
[18] Kelly, Alonzo (1994). A 3D state space formulation of a Numerical Algorithms (Second ed.). Philadelphia, PA:
navigation Kalman lter for autonomous vehicles (PDF). Society for Industrial and Applied Mathematics. p. 680.
DTIC Document: 13. 2006 Corrected Version ISBN 978-0-89871-521-7.
[19] Reid, Ian; Term, Hilary. Estimation II (PDF). www. [33] Masreliez, C. Johan; Martin, R D (1977). Robust
robots.ox.ac.uk. Oxford University. Retrieved 6 August Bayesian estimation for the linear model and robustifying
2014. the Kalman lter. IEEE Transactions on Automatic Con-
trol. 22 (3): 361371. doi:10.1109/TAC.1977.1101538.
[20] Rajamani, Murali (October 2007). Data-based Tech-
[34] Ltkepohl, Helmut (1991). Introduction to Multiple Time
niques to Improve State Estimation in Model Predictive
Series Analysis. Heidelberg: Springer-Verlag Berlin,. p.
Control (PDF) (PhD Thesis). University of Wisconsin
435.
Madison.
[35] Gabriel T. Terejanu (2012-08-04). Discrete Kalman Fil-
[21] Rajamani, Murali R.; Rawlings, James B. (2009).
ter Tutorial (PDF). Retrieved 2016-04-13.
Estimation of the disturbance structure from
data using semidenite programming and opti- [36] Anderson, Brian D. O.; Moore, John B. (1979). Optimal
mal weighting. Automatica. 45 (1): 142148. Filtering. Englewood Clis, NJ: Prentice Hall, Inc. pp.
doi:10.1016/j.automatica.2008.05.032. 176190. ISBN 0-13-638122-7.
[22] Autocovariance Least-Squares Toolbox. Jbr- [37] Rauch, H.E.; Tung, F.; Striebel, C. T. (August
www.che.wisc.edu. Retrieved 2014-06-02. 1965). Maximum likelihood estimates of linear dy-
namic systems. AIAA Journal. 3 (8): 14451450.
[23] Three optimality tests with numerical examples are de- Bibcode:1965AIAAJ...3.1445.. doi:10.2514/3.3166.
scribed in Peter, Matisko, (2012). Optimality Tests and
Adaptive Kalman Filter. 16th IFAC Symposium on Sys- [38] Einicke, G.A. (March 2006). Optimal and Robust Non-
tem Identication. 16th IFAC Symposium on System causal Filter Formulations. IEEE Trans. Signal Process-
Identication. p. 1523. doi:10.3182/20120711-3-BE- ing. 54 (3): 10691077. Bibcode:2006ITSP...54.1069E.
2027.00011. ISBN 978-3-902823-06-9. doi:10.1109/TSP.2005.863042.
17
[39] Einicke, G.A. (April 2007). Asymptotic Opti- quasi-norms. IEEE Transactions on Signal Process-
mality of the Minimum-Variance Fixed-Interval ing. 58 (4): 24052409. Bibcode:2010ITSP...58.2405C.
Smoother. IEEE Trans. Signal Processing. 55 doi:10.1109/TSP.2009.2038959.
(4): 15431547. Bibcode:2007ITSP...55.1543E.
doi:10.1109/TSP.2006.889402. [51] Zachariah, Dave; Chatterjee, Saikat; Jansson, Magnus
(2012). Dynamic Iterative Pursuit. IEEE Trans-
[40] Einicke, G.A.; Ralston, J.C.; Hargrave, C.O.; Reid, D.C.; actions on Signal Processing. 60 (9): 49674972.
Hainsworth, D.W. (December 2008). Longwall Min- doi:10.1109/TSP.2012.2203813.
ing Automation. An Application of Minimum-Variance
Smoothing. IEEE Control Systems Magazine. 28 (6): 28 [52] Vasebi, Amir; Partovibakhsh, Maral; Bathaee, S. Mo-
37. doi:10.1109/MCS.2008.929281. hammad Taghi (2007). A novel combined battery model
for state-of-charge estimation in lead-acid batteries based
[41] Einicke, G.A. (December 2009). Asymptotic Op- on extended Kalman lter for hybrid electric vehicle ap-
timality of the Minimum-Variance Fixed-Interval plications. Journal of Power Sources. 174: 3040.
Smoother. IEEE Trans. Automatic Control. 54 doi:10.1016/j.jpowsour.2007.04.011.
(12): 29042908. Bibcode:2007ITSP...55.1543E.
doi:10.1109/TSP.2006.889402. [53] Vasebi, A.; Bathaee, S.M.T.; Partovibakhsh, M. (2008).
Predicting state of charge of lead-acid batteries for
[42] Einicke, G.A. (December 2014). Iterative hybrid electric vehicles by extended Kalman lter.
Frequency-Weighted Filtering and Smoothing Pro- Energy Conversion and Management. 49: 7582.
cedures. IEEE Signal Processing Letters. 21 doi:10.1016/j.enconman.2007.05.017.
(12): 14671470. Bibcode:2014ISPL...21.1467E.
doi:10.1109/LSP.2014.2341641. [54] Fruhwirth, R. (1987). Application of Kalman ltering
to track and vertex tting. Nucl. Instrum. Meth. A262
[43] Julier, Simon J.; Uhlmann, Jerey K. (1997). A new ex- (23): 444450. Bibcode:1987NIMPA.262..444F.
tension of the Kalman lter to nonlinear systems (PDF). doi:10.1016/0168-9002(87)90887-4.
Int. Symp. Aerospace/Defense Sensing, Simul. and
Controls. Signal Processing, Sensor Fusion, and Target [55] Harvey, Andrew C. (1994). Applications of the Kalman
Recognition VI. 3: 182. Bibcode:1997SPIE.3068..182J. lter in econometrics. In Bewley, Truman. Advances in
doi:10.1117/12.280797. Retrieved 2008-05-03. Econometrics. New York: Cambridge University Press.
pp. 285f. ISBN 0-521-46726-8.
[44] Gustafsson, Fredrik; Hendeby, Gustaf (2012). Some
Relations Between Extended and Unscented Kalman Fil- [56] Boulfelfel, D.; Rangayyan, R.M.; Hahn, L.J.; Kloiber,
ters. IEEE Transactions on Signal Processing. 2: 545 R.; Kuduvalli, G.R. (1994). Two-dimensional restora-
555. tion of single photon emission computed tomography im-
ages using the Kalman lter. IEEE Transactions on Med-
[45] Wan, E.A.; Van Der Merwe, R. (2000). The un- ical Imaging. 13 (1): 102109. doi:10.1109/42.276148.
scented Kalman lter for nonlinear estimation. PMID 18218487.
Proceedings of the IEEE 2000 Adaptive Systems for
Signal Processing, Communications, and Control [57] Bock, Y.; Crowell, B.; Webb, F.; Kedar, S.; Clayton,
Symposium (Cat. No.00EX373) (PDF). p. 153. R.; Miyahara, B. (2008). Fusion of High-Rate GPS and
doi:10.1109/ASSPCC.2000.882463. ISBN 0-7803- Seismic Data: Applications to Early Warning Systems for
5800-7. Mitigation of Geological Hazards. American Geophysi-
cal Union. 43: 01. Bibcode:2008AGUFM.G43B..01B.
[46] Bucy, R.S. and Joseph, P.D., Filtering for Stochastic Pro-
cesses with Applications to Guidance, John Wiley & Sons, [58] Wolpert, D. M.; Miall, R. C. (1996). Forward Mod-
1968; 2nd Edition, AMS Chelsea Publ., 2005. ISBN 0- els for Physiological Motor Control. Neural Netw. 9
8218-3782-6 (8): 12651279. doi:10.1016/S0893-6080(96)00035-4.
PMID 12662535.
[47] Jazwinski, Andrew H., Stochastic processes and ltering
theory, Academic Press, New York, 1970. ISBN 0-12-
381550-9
24 Further reading
[48] Kailath, T. (1968). An innovations approach to least-
squares estimation--Part I: Linear ltering in additive Einicke, G.A. (2012). Smoothing, Filtering and Pre-
white noise. IEEE Transactions on Automatic Control. 13 diction: Estimating the Past, Present and Future. Ri-
(6): 646655. doi:10.1109/TAC.1968.1099025. jeka, Croatia: Intech. ISBN 978-953-307-752-9.
[49] Vaswani, Namrata (2008). Kalman ltered
Jinya Su; Baibing Li; Wen-Hua Chen (2015).
Compressed Sensing. 2008 15th IEEE Interna-
tional Conference on Image Processing. p. 893.
On existence, optimality and asymptotic sta-
doi:10.1109/ICIP.2008.4711899. ISBN 978-1-4244- bility of the Kalman lter with partially ob-
1765-0. served inputs. Automatica. 53: 149154.
doi:10.1016/j.automatica.2014.12.044.
[50] Carmi, Avishy; Gurl, Pini; Kanevsky, Dimitri (2010).
Methods for sparse signal recovery using Kalman l- Gelb, A. (1974). Applied Optimal Estimation. MIT
tering with embedded pseudo-measurement norms and Press.
18 25 EXTERNAL LINKS
Kalman, R.E. (1960). A new approach to Maybeck, Peter S. (1979). Chapter 1. Stochastic
linear ltering and prediction problems (PDF). Models, Estimation, and Control (PDF). Mathemat-
Journal of Basic Engineering. 82 (1): 3545. ics in Science and Engineering. 141-1. New York:
doi:10.1115/1.3662552. Archived from the origi- Academic Press. ISBN 0-12-480701-1.
nal (PDF) on 2008-05-29. Retrieved 2008-05-03.
Moriya, N. (2011). Primer to Kalman Filtering: A
Kalman, R.E.; Bucy, R.S. (1961). New Results in Physicist Perspective. New York: Nova Science Pub-
Linear Filtering and Prediction Theory. lishers, Inc. ISBN 978-1-61668-311-5.
Harvey, A.C. (1990). Forecasting, Structural Time Dunik, J.; Simandl M.; Straka O. (2009). Meth-
Series Models and the Kalman Filter. Cambridge ods for Estimating State and Measurement Noise
University Press. Covariance Matrices: Aspects and Comparison.
15th IFAC Symposium on System Identication,
Roweis, S.; Ghahramani, Z. (1999). A 2009. 15th IFAC Symposium on System Iden-
Unifying Review of Linear Gaussian Mod- tication, 2009. France. pp. 372377.
els. Neural Computation. 11 (2): 305345. doi:10.3182/20090706-3-FR-2004.00061. ISBN
doi:10.1162/089976699300016674. PMID 978-3-902661-47-0.
9950734.
Chui, Charles K.; Chen, Guanrong (2009). Kalman
Simon, D. (2006). Optimal State Estimation: Filtering with Real-Time Applications. Springer Se-
Kalman, H Innity, and Nonlinear Approaches. ries in Information Sciences. 17 (4th ed.). New
Wiley-Interscience. York: Springer. p. 229. ISBN 978-3-540-87848-3.
Stengel, R.F. (1994). Optimal Control and Estima- Spivey, Ben; Hedengren, J. D. and Edgar, T. F.
tion. Dover Publications. ISBN 0-486-68200-5. (2010). Constrained Nonlinear Estimation for In-
dustrial Process Fouling. Industrial & Engineer-
Warwick, K. (1987). Optimal observers ing Chemistry Research. 49 (17): 78247831.
for ARMA models (PDF). International doi:10.1021/ie9018116.
Journal of Control. 46 (5): 14931503.
doi:10.1080/00207178708933989. Retrieved Thomas Kailath; Ali H. Sayed; Babak Hassibi
2008-05-03. (2000). Linear Estimation. NJ: PrenticeHall.
ISBN 978-0-13-022464-4.
Bierman, G.J. (1977). Factorization Methods for
Discrete Sequential Estimation. Mathematics in Sci- Ali H. Sayed (2008). Adaptive Filters. NJ: Wiley.
ence and Engineering. 128. Mineola, N.Y.: Dover ISBN 978-0-470-25388-5.
Publications. ISBN 978-0-486-44981-4.
Bozic, S.M. (1994). Digital and Kalman ltering. 25 External links
ButterworthHeinemann.
Haykin, S. (2002). Adaptive Filter Theory. Prentice A New Approach to Linear Filtering and Prediction
Hall. Problems, by R. E. Kalman, 1960
Liu, W.; Principe, J.C. and Haykin, S. (2010). Ker- How a Kalman lter works, in pictures. Illuminates
nel Adaptive Filtering: A Comprehensive Introduc- the Kalman lter with pictures and colors
tion. John Wiley. KalmanBucy Filter, a derivation of the Kalman
Bucy Filter
Manolakis, D.G. (1999). Statistical and Adaptive
signal processing. Artech House. MIT Video Lecture on the Kalman lter on
YouTube
Welch, Greg; Bishop, Gary (1997). SCAAT:
incremental tracking with incomplete infor- An Introduction to the Kalman Filter, SIGGRAPH
mation (PDF). SIGGRAPH '97 Proceedings 2001 Course, Greg Welch and Gary Bishop
of the 24th annual conference on Computer
Kalman Filter webpage, with lots of links
graphics and interactive techniques. ACM
Press/Addison-Wesley Publishing Co. pp. Kalman lters used in Weather models (PDF).
333344. doi:10.1145/258734.258876. ISBN SIAM News. 36 (8). October 2003.
0-89791-896-7.
Haseltine, Eric L.; Rawlings, James B. (2005).
Jazwinski, Andrew H. (1970). Stochastic Processes Critical Evaluation of Extended Kalman Filter-
and Filtering. Mathematics in Science and Engi- ing and Moving-Horizon Estimation. Industrial
neering. New York: Academic Press. p. 376. ISBN & Engineering Chemistry Research. 44 (8): 2451.
0-12-381550-9. doi:10.1021/ie034308l.
19
26.2 Images
File:Basic_concept_of_Kalman_filtering.svg Source: https://upload.wikimedia.org/wikipedia/commons/a/a5/Basic_concept_of_
Kalman_filtering.svg License: CC0 Contributors: Own work Original artist: Petteri Aimonen
File:Edit-clear.svg Source: https://upload.wikimedia.org/wikipedia/en/f/f2/Edit-clear.svg License: Public domain Contributors: The
Tango! Desktop Project. Original artist:
The people from the Tango! project. And according to the meta-data in the le, specically: Andreas Nilsson, and Jakub Steiner (although
minimally).
File:HMM_Kalman_Filter_Derivation.svg Source: https://upload.wikimedia.org/wikipedia/commons/8/81/HMM_Kalman_Filter_
Derivation.svg License: Public domain Contributors: Own work Original artist: Qef
File:Kalman.png Source: https://upload.wikimedia.org/wikipedia/commons/d/da/Kalman.png License: CC BY-SA 4.0 Contributors:
Own work Original artist: Lipkenszaal
File:Kalman_filter_model_2.svg Source: https://upload.wikimedia.org/wikipedia/commons/a/a0/Kalman_filter_model_2.svg License:
Public domain Contributors: Own work (Original text: I made this image myself, and I release it to the public domain.) Original artist:
User:Headlessplatter
File:Question_book-new.svg Source: https://upload.wikimedia.org/wikipedia/en/9/99/Question_book-new.svg License: Cc-by-sa-3.0
Contributors:
Created from scratch in Adobe Illustrator. Based on Image:Question book.png created by User:Equazcion Original artist:
Tkgd2007
File:Wiki_letter_w_cropped.svg Source: https://upload.wikimedia.org/wikipedia/commons/1/1c/Wiki_letter_w_cropped.svg License:
CC-BY-SA-3.0 Contributors: This le was derived from Wiki letter w.svg: <a href='//commons.wikimedia.org/wiki/File:
Wiki_letter_w.svg' class='image'><img alt='Wiki letter w.svg' src='https://upload.wikimedia.org/wikipedia/commons/thumb/6/6c/Wiki_
letter_w.svg/50px-Wiki_letter_w.svg.png' width='50' height='50' srcset='https://upload.wikimedia.org/wikipedia/commons/thumb/6/6c/
Wiki_letter_w.svg/75px-Wiki_letter_w.svg.png 1.5x, https://upload.wikimedia.org/wikipedia/commons/thumb/6/6c/Wiki_letter_w.svg/
100px-Wiki_letter_w.svg.png 2x' data-le-width='44' data-le-height='44' /></a>
Original artist: Derivative work by Thumperward