Optimization and estimation techniques for passive acoustic source localization. (Techniques d'optimisation et de prédiction dans le contexte de la localisation acoustique passive)

The performance of acoustic passive source localization based on a multiple sensor system does not only depend on the chosen estimation algorithms, but is also strongly correlated to the geometry of the sensor network and the position of the source. This thesis approaches the optimization of the estimation procedure, by utilizing an optimal microphone setup. In order to carry out this optimization procedure for a ”quasi-static” source, three performance measures, the Cramer-Rao Lower Bound (CRLB), the Geometric Dilution of Precision (GDOP) and the condition number, are addressed, compared and evaluated. While the two former describe the influence of measurement noise with known probability density function, the latter is a non-statistical measure. Considering zero-mean Gaussian noise and a linearized model estimator, it is shown that all three approaches lead to the same configuration. The performance measures are extended for a moving source proposing two approaches. The first one is to represent the surveillance area by multiple representative points. In order to assure a good coverage of the zone the average performance measure of all these points is minimized. The second, dynamic approach models the movement of the source using a state-space representation. Recursive Bayesian estimators, such as the Kalman filter for linear systems, predict the most likely upcoming position of the source. Utilizing an adaptive microphone network, only those microphones, which minimize the cost function for this predicted position, are then selected to carry out the estimation procedure. te l-0 04 26 73 2, v er si on 1 27 O ct 2 00 9