Jethro's Braindump

Extended Kalman Filter

Bayes Filter, Kalman Filter, Information Filter

Key Idea

Remove linearity assumption from the Kalman Filter:

\begin{align} x_t &= g(u_t, x_{t-1}) + \epsilon_t \\\
z_t &= h(x_t) + \gamma_t \end{align}

Where function \(g\) and replaces \(A_t, B_t\), and function \(h\) replaces \(C_t\) respectively.

The belief remains approximated by a Gaussian, represented by mean \(\mu_t\) and covariance \(\Sigma_t\). This belief is approximate, unlike in Kalman filters.

Linearization is key to EKFs. EKFs use first-order Taylor expansion for \(g\) to construct a linear approximation to a function \(g\) from its value and slope. The slope is given by the partial derivative:

\begin{equation} g’ (u_t, x_{t-1}) := \frac{\partial g(u_t, x_{t-1})}{\partial x_{t-1}} \end{equation}

Both \(g\) and the slope depend on the argument of \(g\). We choose the most likely argument: the mean of the posterior \(\mu_{t-1}\), giving:

\begin{align} g(u_t, x_{t-1}) \approx g(u_t, \mu_{t-1}) + g’(u_t, \mu_{t-1}) (x_{t-1} - \mu_{t-1}) \end{align}

Where we can define \(g’(u_t, \mu_{t-1}) := G_t\). \(G_t\) is the Jacobian matrix, with dimensions \(n \times n\), where \(n\) is the dimensions of the state.

Similarly, \(h\) is linearized as:

\begin{equation} h(x_t) \approx h(\overline{\mu}_t) + H_t (x_t - \overline{\mu}_t) \end{equation}


\begin{algorithm} \caption{Extended Kalman Filter} \label{ekf} \begin{algorithmic}[1] \Procedure{ExtendedKalmanFilter}{$\mu_{t-1}, \Sigma_{t-1}, \mu_t, \z_t$} \State $\overline{\mu}_t = g(u_t, \mu_{t-1})$ \State $\overline{\Sigma}_t = G_t \Sigma_{t-1} G_t^T + R_t$ \State ${K}_t = \overline{\Sigma}_t H_t^T (H_t \overline{\Sigma}_t H_t^T + Q_t)^{-1}$ \State $\mu_t = \overline{\mu}_t + K_t(z_t - h(\overline{\mu}_t))$ \State $\Sigma_t = (I - K_t H_t) \overline{\Sigma}_t$ \State \Return $\mu_t, \Sigma_t$ \EndProcedure \end{algorithmic} \end{algorithm}


Since the belief is modelled as a multi-variate Gaussian, it is incapable of modelling multimodal beliefs. One extension is to represent posteriors as a mixture of Gaussians. These are called multi-hypothesis Kalman filters.


There are multiple ways for linearization. The unscented Kalman filter probes the function to be linearized at selected points, and calculates a linearized approximation based on the outcomes of the probes. Moments matching linearizes while preserving the true mean and true covariance of the posterior distribution.

Links to this note