Loading...
FCND - 04 - Estimation - 03 - Extended Kalman Filters
Previous:
+
FCND - 04 - Estimation - 02 - Intro to Sensors
Next:
+
FCND - 04 - Estimation - 04 - 3D EKF and UKF
6. Recursive Averages
To avoid memory and performance problems of keeping track of all previous measurements, use the recursive average formula
Markov Assumption
Updated estimate only depends on
current observation
and
previous estimate
Exponential Moving Average
New measurements don’t affect the recursive average after a long period of time t. So we use exponential averaging.
but this also has similar drawbacks to the normal average.
7. Need for Control
We need control to solve the drawbacks of the exponential moving average
We’re going to use our own model of the system to incorporate our control input and expected output.
9. Bayes Filter
If we know the control input we can model what we expect the measurement to be.
We use the model to predict the measurement
We use measurements to correct the prediction
We increase uncertainty in our estimate
Then we make a measurement to decrease the uncertainty
Predict function
bel_bar is an integral, but can be a closed interval.
When the belief is gaussian, it takes the form of matrix operations over the mean and covariance matrix.
Update function
We use the predicted state and update it with a new estimation using Bayes’ Rule. It uses P(z|x) x belief
11. Kalman Filter Math
State Transition Function
(g)
You’ll usually see Kalman Filter math expressed in the matrix notation below.
This is equivalent to the more readable version of it above, which just says that to advance the state:
State
Please turn on JavaScript to use Paper in all of its awesomeness. ^_^
6. Recursive Averages
Markov Assumption
Exponential Moving Average
7. Need for Control
9. Bayes Filter
Predict function
Update function
11. Kalman Filter Math
State Transition Function (g)