Gedistribueerde Geluidsbronlokalisatie in Draadloze Akoestische Sensornetwerken

Sound source localization (SSL) is a fundamental task in acoustic signal processing, with applications including teleconferencing, human-machine interaction, spatial audio analysis, surveillance, and health monitoring. Traditionally, sound source localization systems rely on compact microphone arrays connected to a centralized processing unit. While such systems can achieve high localization accuracy, they are typically limited in spatial coverage and flexibility. In recent years, wireless acoustic sensor networks (WASNs) consisting of spatially distributed microphones have emerged as a promising alternative for large-scale acoustic sensing. By allowing multiple devices equipped with microphones to cooperate through wireless communication, WASNs enable flexible deployment, scalability, and improved robustness to node failures. However, their distributed nature introduces significant challenges, including limited communication bandwidth, constrained computational resources, and the need for decentralized processing. This dissertation investigates how SSL algorithms can be redesigned to operate efficiently in such distributed environments while maintaining robustness under noisy and reverberant conditions. The focus is on steered response power (SRP)-based localization methods, which are widely used due to their strong robustness to noise and reverberation. Conventional SRP formulations, however, typically require centralized processing and exhaustive use of microphone pairs, resulting in high communication and computational costs that limit their applicability in WASNs. To address these limitations, this thesis develops several distributed and communication-efficient extensions of the SRP framework. First, a comprehensive tutorial review of SRP-based sound source localization methods is presented. This review analyzes more than two hundred publications on SRP and its variants and categorizes existing approaches according to their objectives, including computational complexity reduction, robustness improvements in adverse acoustic environments, and extensions to multi-source localization. The review provides a structured overview of the SRP literature and establishes the methodological foundation for the developments presented in this dissertation. Second, a distributed SRP framework for WASNs is introduced. In the proposed approach, each node computes a local SRP map using only the microphone pairs available within that node, thereby avoiding the transmission of raw microphone signals across the network. Two fusion strategies are investigated: a centralized approach in which local SRP maps are aggregated in a fusion center, and a fully decentralized approach in which nodes iteratively exchange and average their maps with neighboring nodes. Simulation results demonstrate that these strategies achieve localization performance comparable to centralized SRP methods while significantly reducing communication requirements. Third, the thesis addresses the high communication and computational cost associated with large distributed microphone arrays through microphone pair selection strategies. Three methods are proposed to identify informative microphone pairs for SRP-based localization. The first method selects pairs exhibiting strong cross-correlation peaks at zero lag, while the second and third methods exploit sparsity properties of the cross-correlation function. To enable fully distributed operation, cross-correlations are approximated locally and a gossip-based algorithm is used to perform decentralized top-K pair selection. Experimental results show that accurate localization can be achieved using only a subset of microphone pairs, substantially reducing communication and computational requirements. Finally, an interpretable SRP formulation is proposed for multi-source localization. Instead of evaluating the SRP function on a dense spatial grid, the proposed method evaluates the response only at candidate locations obtained from intersections of hyperbolic time-difference-of-arrival (TDOA) loci derived from dominant generalized cross-correlation with phase transform (GCC-PHAT) peaks. The resulting representation is factorized using an overcomplete nonnegative matrix factorization (NMF) model to estimate multiple source locations. Simulation studies demonstrate improved localization performance compared to classical SRP peak picking and low-rank NMF approaches, particularly in noisy and reverberant environments. Together, these contributions demonstrate that SRP-based sound source localization can be effectively adapted to distributed scenarios involving WASNs.