Nonlinear Observation Models - motoq/bierman GitHub Wiki

Slight modifications have been made to the algorithms in Bierman's book to better address the nonlinear nature of the primary example problem used for this project. Instead of generating the observation residual using the partials of the observations w.r.t. the solve for parameters,

delta = z - Ax

the residual is formed using the (nonlinear) observation model with all parameters updated based on the most recent estimate:

delta = z - f(x)

Likewise, when estimating a dynamic state, the model is used for the prediction (state propagation) step and not the (linear) state transition matrix. In addition, the dynamic model always makes use of the most recently estimated state vector.

The concept of a linearized extended Kalman filter has essentially been applied to the algorithms in this book. The majority of the examples focus on estimate covariance decomposition and SRIF methods although a stabilized form of the Kalman filter is used for sanity checks and performance comparisons.