Professional Documents
Culture Documents
Nonlinear Filtering - Extended Kalman Filter - Linearization and Random Variables
Nonlinear Filtering - Extended Kalman Filter - Linearization and Random Variables
Winter 2005-06
Nonlinear ltering
81
Nonlinear ltering
nonlinear Markov model: x(t + 1) = f (x(t), w(t)), y(t) = g(x(t), v(t))
f is (possibly nonlinear) dynamics function g is (possibly nonlinear) measurement or output function w(0), w(1), . . . , v(0), v(1), . . . are independent even if w, v Gaussian, x and y need not be
nonlinear ltering problem: nd, e.g., x(t|t1) = E(x(t)|y(0), . . . , y(t1)), x(t|t) = E(x(t)|y(0), . . . , y(t))
83
(x(i))
i=1 (i) mc (i) mc T
(x ) y
(x ) y
where x(1), . . . , x(N ) are samples from the distribution of x, and N is large another method: use Monte Carlo formulas, with a small number of nonrandom samples chosen as typical, e.g., the 90% condence ellipsoid semi-axis endpoints x(i) = x vi, x = V V T
85
Example
x N (0, 1), y = exp(x) (for this case we can compute mean and variance of y exactly)
exact values linearization Monte Carlo (N = 10) Monte Carlo (N = 100) Sigma points (x = x, x 1.5x)
e1/2
86
Ct|t1C + V
1
...
Ct|t1
87
time update linearize dynamics function at x = x(t|t): A = W = f ((t|t), 0) x x f f ((t|t), 0)w ((t|t), 0)T x x w w
time update based on linearization x(t + 1|t) = f ((t|t), 0), x t+1|t = At|tAT + W
replacing linearization with sigma-point estimates yields unscented Kalman lter (UKF)
88
Example
p(t), u(t) R2 are position and velocity of vehicle, with (p(0), u(0)) N (0, I) vehicle dynamics: p(t + 1) = p(t) + 0.1u(t), w(t) are IID N (0, I) measurements: noisy measurements of distance to 9 points pi R2 yi(t) = p(t) pi + vi(t), vi(t) are IID N (0, 0.32)
The Extended Kalman lter 89
u(t + 1) =
u(t) + w(t)
i = 1, . . . , 9,
EKF results
EKF initialized with x(0| 1) = 0, (0| 1) = I, where x = (p, u) pi shown as stars; p(t) as dotted curve; p(t|t) as solid curve
9 8 7 6 5
p2 PSfrag replacements
4 3 2 1 0 1 3
p1
The Extended Kalman lter 810
PSfrag replacements
20
40
60
80
100
120
140
160
180
200
811
1/2
versus t
1.4
1.2
0.8
0.6
PSfrag replacements
0.4
0.2
100
200
300
400
500
600
700
800
900
1000
812