Indoor Location Fingerprinting Based on Data Reduction Dragan Kukolj Faculty of Technical Science, Univ. of Novi Sad, Serbia dragan.kukolj@rt-rk.com Marina Vuckovic RT-RK Computer Based Systems, Novi Sad, Serbia marina.vuckovic@rt-rk.com Szilveszter Pletl Univ. of Szeged, Department of Informatics, Hungary pletl@inf.u-szeged.hu Abstract— Agent localization in indoor wireless environments is a challenging issue. Numerous techniques have been developed. Location fingerprinting, which is based on received signal strength measurements, is a frequently used approach for indoor applications. In this paper, we examine the possibility to obtain the location fingerprinting method characterized with more accurate mapping between the signal- space and the physical-space. An implemented well-known weighted k-nearest neighbor (WkNN) method is enhanced by two steps: a) pre-processing by the unsupervised learning technique during radio map building and b) post-processing of initial estimates obtained by the WkNN localization method. In this post-processing step signal-space and physical-space are transformed and mapped using two techniques of the dimension reduction: principal component analysis and multidimensional scaling. The aim of this transformation step is to de-correlate and refine initially obtained location estimates. Parameters such as number of access points and number of nearest reference nodes are examined for their impact on accuracy of the presented localization techniques. Performances are examined and verified through the experiments with real environment data. Keywords—Location Fingerprinting; Received Signal Strength; Topology-Preserving Mapping; Data Reduction. I. INTRODUCTION Indoor localization techniques have different and numerous applications in the wireless network technology for the positioning and tracking of people or objects. The most popular indoor localization technique is fingerprint- based localization, having the major advantage to exploit already existing wireless network infrastructures, like IEEE 802.15 or IEEE 802.11, which avoids additional deployment costs. The location fingerprinting (LF) is a method of predicting location on basis of pre-recorded measurements of received signal strength (RSS) in a deployment area and inspection of currently measured RSS at the tracked agent. Depending on how the database of pre-recorded RSSs is formed and processed, fingerprinting localization approaches are grouped into deterministic [1] and probabilistic methods [2]. Several techniques that are able to adapt to variations of signal strength over different time periods have been proposed [3], [4]. Fang and Yin [5] shows that, the projection of measured RSS into a de-correlated signal space could improve localization accuracy. In the paper [6] Fang and Yin presented a localization algorithm named discriminated-adaptive neural network able to create nonlinear relationship between RSS and the position. The work of Ni et al. [7] describes enhancement of the WkNN in the indoor environment applying multidimensional scaling (MDS) at pairwise squared distances between RFID tags and the reference points. An approximated RFID tags’ distribution is refined by the subsequent use of Procrustes analysis in order to obtain LF estimation of higher accuracy. The LF approach consists of two phases: an off-line training phase and on-line localization phase. In the off-line phase, RSSs are collected at predetermined positions of the reference points to build the database called a radio map. During the on-line localization phase the localization algorithm estimates the unknown position of tracked wireless sensor using real-time measured RSS samples, the stored radio map database and learned estimation model. The basic idea of this work is to show how accuracy of the classical Euclidean based weighted k - nearest neighbors algorithm applied in LF for an indoor environment can be improved by inserting two independent algorithmic steps: 1) data pre-processing during training phase of LF process using an unsupervised learning algorithms; and 2) the linear transform of the distances / similarities between the appropriate RSS vectors and spatial distances. In the pre-processing step two alternative unsupervised techniques characterized by the common topology- preserving feature are used. These techniques are: the Neural Gas (NG) algorithm [8], [9], and Re-Organizing Neural Network (RONN) [10], [11]. The aim of this step is 2011 International Conference on Broadband and Wireless Computing, Communication and Applications 978-0-7695-4532-5/11 $26.00 © 2011 IEEE DOI 10.1109/BWCCA.2011.52 327