Abstract

Audio-based Time Difference of Arrival (TDOA) fingerprinting systems determine positions by analysing sound signal delays captured by multiple microphones. Their performance depends on the density of the TDOA fingerprint database, where higher density enhances localization accuracy but also extends localization time. To address this trade-off, clustering is applied to the fingerprint database, where the choice of clustering algorithm plays a critical role in performance. Furthermore, the fingerprint similarity metric and the microphone configuration used to create the database also significantly influence the system's effectiveness. This paper investigated the clustering performance of two widely used algorithms, k-means and k-medoids, for application in an audio-based TDOA fingerprinting system. The analysis considered varying numbers of clusters (K = 2 to 7), different fingerprint similarity metrics (Euclidean and Manhattan distances), and varying microphone configurations (square and rectangular). Using silhouette scores as a clustering performance metric, the results indicated that employing a rectangular microphone configuration to create the fingerprint database yielded better clustering performance for both algorithms. Moreover, using Euclidean distance as the fingerprint similarity metric resulted in the formation of well-defined clusters by both algorithms. Although both clustering algorithms demonstrated comparable performance, k-means typically outperformed, particularly when the number of clusters was smaller.

Key words: Clustering algorithms, Database density, k-means, k-medoids, TDOA-based fingerprinting systems