Search space reduction in DCM positioning using unsupervised clustering

Abstract

Database correlation methods (DCM) are used to locate mobile stations (MS's) in wireless networks. A target radio-frequency (RF) fingerprint - measured by the target mobile station - is compared with georeferenced RF fingerprints, previously stored in a correlation database (CDB). In this paper, two unsupervised clustering techniques (K-medians and Kohonen Layer) were applied to reduce the search space inside the CDB. The clustering effects on the computational cost of the positioning method and on the positioning accuracy were experimentally evaluated using 46200 target fingerprints and a CDB with 924 reference fingerprints, containing Received Signal Strength (RSS) values of 136 WiFi 802.11b/g networks in a 12-floor building. A reduction of 81% in the average time to produce a position fix was observed, as well as a 38% decrease in the DCM average positioning error and a 6% improvement in the floor identification accuracy.