Sound Source Localization in a 3D Multi-Surface Environment Using Learning-Based Methods

This study presents a three-dimensional localization of a ball drop in a multi-surface environment using cost-effective data acquisition devices, and proposes two learning-based methods to improve the baseline classical localization method based on GCC-PHAT-estimated TDOA values. The first proposed method employs a ball drop surface classification executed by Random Forest and XGBoost models to enhance localization performance, whereas the second method conducts region classification by using both basic and multi-branch neural network models, achieving improved and more efficient localization performance. A dataset of real-world sparse ball drop samples was collected in a controlled multi-surface environment for experimental execution, where results demonstrate that the two proposed methods improve the localization performance across various scenarios, with the region classification method significantly outperforming the baseline method, achieving a 17% reduction in the Mean Euclidean Distance error metric. Furthermore, the surface and region classification methods reduce the localization time compared to the baseline method by 50% and 97%, respectively. The obtained results reveal the potential of combining simple learning-based models with affordable microphone sensors for achieving an accurate localization of sparse audio samples in a multi-surface environment.