MDS-based Localization Algorithm for RFID Systems Wenbo Shi and Vincent W.S. Wong Department of Electrical and Computer Engineering, The University of British Columbia, Vancouver, Canada E-mail: {wenbos, vincentw}@ece.ubc.ca Abstract— In radio frequency identification (RFID) systems, location information is of great importance to provide location- aware services combined with identification. Conventional RFID systems can only provide coarse localization information. In this paper, we propose a novel approach named MDS-RFID to locate active RFID tags based on multidimensional scaling (MDS), an efficient data analysis technique. The approach has the advantage of fully utilizing the distance information in the network simultaneously, and thus can achieve better localization results than previous multilateration-based methods. The MDS- RFID algorithm first infers the tag-to-reader distances from the received signal strength (RSS). To obtain the distance matrix, the inter-tag distances are estimated using a triangular method. Then, classical MDS algorithms can be applied to determine the estimated locations of the tags. An optional refinement step can be added to further improve the accuracy using maximum likelihood estimation at the expense of additional computation costs. Simulation results show that the MDS-RFID algorithm can achieve a significant gain in accuracy over the previous localization schemes based on multilateration using only a few readers. I. I NTRODUCTION Radio frequency identification (RFID) is a wireless tech- nology that enables data communications via radio waves between a reader and a tag which is attached to an object for identification and tracking [1]. An RFID system is composed of RFID tags, readers, and servers. An RFID tag is a micro- chip combined with an antenna. It is able to store the object information that can be retrieved by the reader via radio waves, which provides unique identification for every object in the system.The reader has an interface to communicate with the server, sending the data it receives from the tags to the server via wired or wireless communications. RFID systems can be enhanced greatly if the identification information is combined with locations to provide many attracting location- aware applications (e.g., mobility control, resource allocation, security, and service discovery). Several RFID localization schemes have been proposed in the literature. They can be classified into: distance estimation, scene analysis, and proximity [2]. The algorithms based on distance estimation use multilateration to estimate the posi- tions of RFID tags. The ranging techniques include received signal strength (RSS) [3], time of arrival (TOA) [4], time difference of arrival (TDOA) [5], and received signal phase (RSP) [6]. Scenes analysis approaches use k-nearest-neighbor (kNN) [7] or probabilistic methods [8], [9] to locate RFID tags by matching the measurements to the fingerprints collected from the environment. Proximity is another group of methods to locate RFID tags assuming that the position of the target tag is the same with the reader which detects the tag. Based on the type of RFID tags used in the system, the localization schemes can be classified as being active [3], [7] or passive [10], [11] RFID tag localization as well. An active RFID tag powered by a cell battery has a much longer communications range than a passive tag. In this work, we mainly focus on active RFID tag localization. SpotON [3] is a RFID location sensing system which belongs to the family of distance estimation. SpotON uses RSS measurements from long range active RFID tags to approximate the distances between tags and readers. Multiple readers collect the RSS measurements and then transform the RSS measurements into distance estimations through a function defined with empirical data. Multilateration is then applied to calculate the positions of the tags. LANDMARC [7] is an active RFID locating system using scene analysis. It requires reference tags to be deployed at known positions. RSS measurements from multiple readers are used to locate the target RFID tag using the locations of the k nearest reference tags. Different weights are imposed on the k nearest reference tags to estimate the location of the target tag. In the LANDMARC system, dense deployment of reference tags are necessary for good performances. Several extensions based on the LANDMARC system have also been proposed in the literature [12], [13]. In the context of localization, multidimensional scaling (MDS) [14] can be used to solve the localization problem efficiently where the similarity measures in MDS corre- spond to the Euclidean distances. MDS-based localization algorithm (MDS-MAP) has been well studied in wireless sensor networks [15]–[17]. The MDS-MAP algorithm first starts with forming the distance matrix by running a shortest- path algorithm. Next, MDS is performed to determine the estimated locations of the sensors that best fit those dis- tance measurements. Finally, anchors are used to calculate the absolute coordinates based on linear transformations and rotations. MDS-MAP is a popular localization algorithm in wireless sensor networks and it has many variants (e.g., MDS- MAP(C) [15] which uses metric MDS, MDS-MAP(O) [18] which adopts ordinal MDS, and MDS-MAP(P) [17] which is a distributed scheme using patches of relative map). Additional