Research on Sound Source Localization Based on Random Forest

To address the issue in location fingerprinting methods where the Euclidean distance between data points tends to equalize as feature dimensionality increases, leading to a loss of metric differentiation, a random forest-based localization method is proposed. The method involves two stages: an offline sampling stage and an online localization stage. During the offline stage, sound source signals are captured at each localization reference point, and time-delay features are extracted to construct a localization database using the extracted features and reference point location information. In the online stage, a time delay estimation algorithm combining the initial segment of the envelope with cross-correlation (ISE-CC) is used to extract the time-delay features of the signal at the target point, and localization is achieved using random forest (RF) predictions. Simulation experiments comparing the proposed method with nearest neighbor (NN) and weighted k-nearest neighbor (WKNN) demonstrate that the proposed method achieves minimal deviation in sound source position estimation and satisfies the requirements for practical applications.