Bayesian Estimation and Tracking: A Practical Guide

17.3 Performance Comparison of All DIFAR Tracking Filters

In this section, we present an analysis of the performance of many of the tracking filters applied to the DIFAR case study. We have delayed this performance analysis until all of the filters have been presented in Parts II and III. Since all of the results presented so far have been based on simulated observation data, the best method of assessing performance is by comparison of the x-and y-axis position RMSE for the various filters. For each signal SNR, a RMS position error comparison plot is presented for four of the Gaussian filters (EKF, UKF, SSKF, and GHKF) and three of the particle filters (BPF, APF, and GSPF). The acronym GSPF represents a combination particle filter with an UKF used as the importance density. For all particle filters, 3000 Monte Carlo samples were used.

As a first example, Figure 17.5 shows a comparison of both the x-and y-axis RMS position errors as a function of the true position of the target ship when the signal SNR is 20 dB. One can immediately see that the BPF RMS position errors grow quickly due to the three divergent tracks (out of 100 Monte Carlo runs) at 20 dB SNR, as can be seen in Figure 16.5. This is just another illustration of the problems associated with initialization sensitivity for the BPF. Since all of the filters were initialized with the same set of initialization values, this additional tuning needed for the BPF is unacceptable for application to real-world tracking.

Figure ...