Prateek Tandon
Generic Problem
Imagine watching a small bird flying
through a dense jungle.
You glimpse intermittent flashes of
motion.
You want to guess where the bird is and
where it may be in the next time step.
Birds state might be 6-dimensional:
[x,y,z,x,y,z] three variables for position
and three for velocity.
Kalman Filter
Xk = Fk xk-1 + Bk uk + wk (state update)
Zk = Hkxk + vk (measurement update)
Fk state transition
X current state model
X last state
Bk control input
U control input
W ~ N(0,Q ), represents
process noise distributed via
model
multivariate zero-mean normal distribution with
covariance Q
Hk observation
V ~ N(0,R ), represents observation nose distributed via
multivariate zero-mean
normal distribution with
model
covariance R
k
k-1
k
UPDATE:
Predicted
Covariance
Innovation and Measurement
Residual
Innovation on Covariance
Optimal Kalman Gain
Updated State Estimate
Updated Covariance Estimate
Applications
Radar tracking of
planes/missles/navigation
Smoothing time series data
Stock market
People tracking / hand tracking / etc
Sensor Data
R0
P(R
1)
0.7
0.3
0.7
Rain0
Rain1
R1
P(U
1)
0.9
0.2
Umbrella1
(a) Propagate
Raint+1
Raint+1
(b) Weight,
[Not Umbrella
observed.]
Raint+1
(c) Resample
References
"Kalman Filter." . WIKIPEDIA, 13
APRIL 2013. Web. 13 Apr 2013.
<http://en.wikipedia.org/wiki/Kalman
_filter>.
Russell, Stuart, and Peter Norvig.
Artificial Intelligence: A Modern
Approach. 3rd. New Jersey: Pearson
Education Inc., 2010. Print.