2D cartesian sound source localization in an indoor reverberant environment using deep learning

Inadequate auditory capabilities hinder the development of truly intelligent physical systems. Robots without proficient auditory functions often struggle in unpredictable situations and are largely ineffective in human collaboration and interaction. Sound perception often surpasses visual abilities, especially in dark or cluttered environments. This exceptional feature of auditory perception can be pivotal: aiding disaster robots in locating victims, assisting self-driving cars in avoiding obscured obstacles or pedestrians, and facilitating seamless human-robot interactions even through physical barriers. For accurate sound source pinpointing in real-world scenarios, it is essential to consider the complexity posed by sound wave propagation in such a complex environment. Tackling these issues, this study introduces an innovative method for two-dimensional (2D) sound source localization in Cartesian coordinate system in cluttered, real world indoor settings. This novel method leverages the capabilities of both conventional method and deep learning method by utilizing sound signal combined with environment maps generated by robotic SLAM for the first time, drawing upon information from incoming sound signals and environmental geometry. Ultimately, this research novel method can predict the location of a sound source in 2D Cartesian coordinate with just using small amount of training data faster and easier and more precise due to using sound signal and geometry of the environment. The capability of this research method is evaluated and compared to other state-of-the-art methods at the end. This groundwork paves the way for subsequent studies, adapting the strategy for spaces with multiple sound sources and intricate indoor areas where humans and robots might cooperate in separate rooms.

Author

• Abhari, Maziar

Identifier

hdl:10057/27822

OAI identifier oai:identifier

oai:soar.wichita.edu:10057/27822