-
Choose between a linear, an extended and an unscented Kalman filter, using a camera’s view of a pendulum. See what the library needs from you, how it compares with VQF on recorded motion, and what an update costs on a microcontroller.
-
Build the library, run a filter which tracks a swinging pendulum from camera frames, and match each line it prints to a replay of the run.
-
Choose the fields of a state and see its uncertainty as an ellipse with sigma points on it. Compare the three integrators on one prediction, and set the process noise for a gap between readings.
-
Follow one camera frame from predicted reading to correction. Combine a camera and a gyroscope which report at different times, and describe a sensor error which has a direction.
-
Check one update by hand, and see a covariance and its square-root factor as the same ellipse. Keep predicting while the camera is hidden, and know what a failed step leaves behind.
-
Calibrate the camera from the rod clamped at known angles, and see why one angle is not enough. Then refuse a reading which the filter finds too surprising.
-
Blow on the pendulum with a fan the model does not know about, and give the filter a state to explain it. Then move the sigma points and watch the predicted reading change.