Direct Virtual Sensor design for radar localization systems

In recent decades, a large amount of studies have been focusing on target tracking using radar measurements. Radar systems typically measure the position of the target in polar coordinates, but in many applications it is of interest to obtain the target Cartesian coordinates. However, due to measurement noise, the conversion from polar to Cartesian coordinates using the classical trigonometric formulas may lead to inaccurate results. This problem can be mitigated by using an observer/filter, which can attenuate the noise effects and provide a more accurate estimate of the target Cartesian coordinates. Designing a state observer, the main two following phases have to be expected: first identifying a system model from an experimental data set, and then designing an observer based on the identified model. This two-step approach might be problematic due to some nonlinearities in the identified model. So, a one-step approach is preferable, since the data set is used to direct design the filter and not for identification of the model. The aim of this thesis is to design both the two-step and the one-step procedures in order to highlight the differences between the two and to put an accent to the solution that gets better performances, based on simulations. Regarding the first method, an Extended Kalman Filter (EKF) has been designed in Simulink, while in the one-step procedure a Direct Virtual Sensor (DVS) is engineered using Simulink, the System Identification Toolbox and neural networks. Through the simulations, the results got in terms of root-mean-square error, do give evidence that the proposed one-step procedure may be better than the two-faces described above.