Parametric Models and Algorithms for Direction-of-Arrival Estimation

Abstract

The ability to selectively focus on desired sounds in noisy environments poses a significant challenge with broad applications, including smart devices, driver assistance systems, smart homes, video conferencing, drones, and hearing aids. Acoustic source localization involves identifying the position of a sound source amidst various factors like reflections, reverberation, and background noise. While extensively studied, acoustic source localization remains an active area of research due to its diverse applications. Existing localization algorithms face several challenges that limit their effectiveness and practicality. These challenges include reliance on narrowband models, computational efficiency, adaptability to non-stationary targets, robustness against noise and reverberations, high-resolution localization, and distinguishing between correlated sources. Overcoming these challenges is crucial for the development of advanced localization algorithms that enhance accuracy, efficiency, and reliability in practical scenarios. This thesis is divided into two main parts. Firstly, a comprehensive performance analysis is conducted to evaluate various localization algorithms using real-world datasets, aiming to gain a deep understanding of their capabilities. Secondly, a novel technique called trajectory localization (TL) is proposed, which enables accurate estimation of complex trajectories of multiple moving sources simultaneously, eliminating the need for tracking filters. The technical contributions of this thesis include experimental validation of existing localization algorithms and the development of wideband signal models and algorithms on real-world recordings. Deep learning architecture is introduced that incorporates direction of arrival (DOA) derivatives for improving the temporal continuity of DOA, hence resulting in smoother source trajectories. Next, we develop parametric models and algorithms for joint localization and tracking tasks and explore various trajectory localization algorithms. The effectiveness of the proposed algorithms is demonstrated through their application to real-world recordings in challenging scenarios. Moreover, the proposed models and algorithms have the potential to extend beyond sound waves and be applied to other data types, such as radio waves, expanding their impact across various applications