A bootstrap approach for evaluating source localization performance on real sensor array data
Bootstrap methods for evaluating the performance of source localization methods on real sensor array data without precise a priori knowledge of true source positions and the underlying data distribution are examined. Bootstrap resampling methods are used to measure the standard deviation of bearing estimates achieved by performing delay-and-sum, minimum variance, and MUSIC spatial spectral estimation on a real narrowband towed-array data set. Some simple theoretical guidelines are given to indicate the sample size required to achieve accurate performance estimation. Results suggest that quantitative bearing estimation performance comparisons can be made within the observation time constraints imposed by source and field dynamics. Comparisons of bootstrap estimates to Monte Carlo and theoretical benchmarks provide a means of validating the parametric models assumed by these conventional analysis techniques.