When the sensor is not linear: zoom in until the curve looks straight, use the Jacobian, and track a robot from distance-only radars. 6 short slides, then worked examples and problems. Each slide has an everyday picture, a worked example and Python you can run right on the page.
Free preview: most of every slide is yours to read, with its Python lab. Members get the full text, every example with its solution, and five extra problems per lesson.
Lesson 4Nonlinear sensorsNonlinear sensors: when z = Hx breaksOpen slide →01 · Lesson 4 · Nonlinear sensorsA radar measures distance, not position. Why no matrix H can do that, and what happens to a Gaussian pushed through a curve: it bends and its mean shifts.
Free preview
Lesson 4The JacobianThe Jacobian: the slopes of a curve, seen up closeOpen slide →02 · Lesson 4 · The JacobianZoom in until the curve looks straight. The Jacobian as a row of slopes, the derivation of ∂r/∂x = x/r, and the (3, 4) example with slopes 0.6 and 0.8.
Free preview
Lesson 4Uncertainty in, outUncertainty through a Jacobian: σ² ≈ HJ P HJᵀOpen slide →03 · Lesson 4 · Uncertainty through a JacobianHow position uncertainty becomes range uncertainty: σ_r² ≈ H P Hᵀ. The worked example at (3, 4): 5 m ± 0.71 m, checked against samples.
Free preview
Lesson 4The EKFThe extended Kalman filter: two changes to the updateOpen slide →04 · Lesson 4 · The extended Kalman filterThe EKF is the linear Kalman filter with two changes: the innovation uses the real h(s), and the Jacobian replaces H. One full radar update with real numbers.
Free previewTwo radars that measure only distance track a turning robot. Why one radar is ambiguous, the full EKF in Python, tuning R, and when an EKF fails.
Free preview
Lesson 4RecapRecap: the extended Kalman filterOpen slide →06 · Recap · ReadingThe EKF in one page: nonlinear sensors, the Jacobian, uncertainty through H P Hᵀ, the two changes to the Kalman filter and range-only tracking.