As the quadrocopter, as a system, includes nonlinearities, it is
common practice to employ the Extended Kalman filter (illustrated
in Fig. 6), where a linear approximation is only used for solving
the Riccati equation, a result of which is the Kalman gain. The full,
nonlinear model is used in the propagation of the estimate and in
computing the predicted sensor outputs [34]. This would introduce
a heavy load on the on-board, high-level processor and thus was
not selected for our application. Therefore, a nonlinear part of
the quadrocopter system (mainly the nonlinear coordinate system
transformation from the quadrocopter’s coordinate system K to
the target coordinate system T ) was decoupled from the linear
part of the system and replicated on the path of the acceleration
measurement vector aK entering the Kalman filter (Fig. 7) in order
to enable aT to enter the Kalman filter directly. This enabled us to
employ a basic (linear) form of the Kalman filter that presents a
much lighter load to the processor.