Mathematical modelling and optimization strategies for acoustic source localization in reverberant environments

Abstract

This thesis deals with the problem of indoor acoustic source localization using modern optimization strategies. It includes modeling, algorithms, and calibration, which allows using localization algorithms even when the geometry of the microphones is unknown. The aim of this thesis is to localize robustly and accurately speakers within a reverberant environment equipped with array of microphones. The previous exiting techniques usually required a high number of microphones in order to get high accuracy. During this thesis, we have develop a new method which improves up to 30% the localization accuracy with a reduced number of microphones. Using a low number of microphones is important since it directly reduce the cost and improve the versatility of the final system. On the other hand, we have performed a exhaustive analysis about the PHAT filtering (broadly used in acoustic localization), including all the phenomena involved in acquisition and signal processing. Our analysis improves the knowledge about PHAT filtering, modeling the main aspects involved in acoustic localization. Previous model has yielded a sparse representation of the acoustic source localization scenario. This kind of representation has been demonstrated very convenient for localization since it allows to deal with multiple simultaneous sources easily. Additionally, we have proposed a method for the calibration of pairwise distance using the diffuse noise present in a silent room. The new algorithm is related with previous methods based in coherence. Nevertheless, applying the developed model for PHAT filtering we have been able to introduce physical constraints based on the maximum expected distance between microphones. It allows to improve accuracy and reducing the computational cost. Finally but not least, we have characterize TDOA matrices. We have propose several methods to robust denoise TDOA measurements exploiting low-rank properties of TDOA matrices. Therefore, these methods are not limited to acoustic source localization, but are useful for other techniques such as self-calibration and beamforming, and other technologies (e.g. radar, ultrasound).