Sound Source Localization Using Multi-Dictionary Orthogonal Matching Pursuit in Reverberant Environments

Sound source localization in reverberant environments remains a challenging problem, particularly when precise position estimation is required. Existing DOA estimation methods, while effective in determining sound direction, often fall short of accurate position estimation in adverse acoustic conditions due to their inherent constraints. In this paper, we propose Multi-Dictionary Orthogonal Matching Pursuit (MD-OMP), extending sparse recovery methods by jointly exploiting spatial, temporal, and spectral information for source position estimation. MD-OMP iteratively selects candidate grid entries that maximize joint correlation across time frames and frequency bins, thereby enhancing localization accuracy under reverberant conditions. We also introduce a generalized extension of the algorithm to accommodate an arbitrary number of multidimensional domains, thereby supporting advanced applications such as higher-order source modeling and scenarios with dominant frequency components. To validate the effectiveness of MD-OMP, we arrange microphones along the room walls to form a perimeter array and conduct both extensive simulations and a practical experiment in a conference room setting. Numerical simulations under various acoustic conditions-including different reverberation times, grid densities, frequency ranges, room dimensions, and source types-show that MD-OMP consistently outperforms group-sparsity-based and TDOA-based benchmark methods in localization accuracy. Experimental results in a realistic office environment further verify its practical applicability. Overall, MD-OMP achieves accurate and robust source position estimation across diverse and challenging scenarios.