Abstract:
Within the context of fingerprint database clustering, the density-based spatial clustering of applications with noise (DBSCAN) is notable for its robustness to outliers and ability to handle clusters of different sizes and shapes. However, its high computational burden limits its scalability for dense fingerprint databases. A hybrid two-stage clustering method, the CAP-DBSCAN algorithm, is proposed in this paper, designed to accelerate DBSCAN clustering while ensuring accuracy for fingerprint-based localisation systems. The CAP-DBSCAN algorithm employs the closest access point (CAP) algorithm to pre-cluster the database, while the DBSCAN algorithm performs clustering refinement. It dynamically adjusts the neighborhood radius (Eps) value for each pre-cluster using the k-distance plot method. The performance of the CAP-DBSCAN algorithm is determined across four publicly available received signal strength (RSS)-based fingerprint databases with Euclidean and Manhattan distances as fingerprint similarity metrics. This is benchmarked against the performances of the standard DBSCAN (s-DBSCAN) and k-means++-DBSCAN (k-DBSCAN) algorithms presented in previous research. Simulation results show that the CAP-DBSCAN algorithm consistently outperforms both the s-DBSCAN and k-DBSCAN algorithms, achieving higher silhouette scores, which indicates the generation of more compact and well-defined clusters. Furthermore, the CAP-DBSCAN algorithm demonstrates superior computational efficiency as a result of the CAP algorithm generating well-structured pre-clusters better than those generated by the k-means++ algorithm. This significantly reduces the computational burden of the cluster refinement process. Overall, using Manhattan distance as a fingerprint similarity metric results in the best clustering performance of the CAP-DBSCAN algorithm. These findings underscore the potential of the CAP-DBSCAN algorithm for practical applications in resource-constrained fingerprint-based localization systems. © 2025, Ital Publication. All rights reserved.
