Indoor Positioning via WLAN Channel State Information and Machine Learning Classification Approaches

Abstract

Location-based services and navigation has become ubiquitous in a mobile computing era. An example of a popular navigation system is GPS [1], which uses a trilateration technique but has the limitation of working in outdoor environments. Because of this limitation, the pursuit for an indoor positioning system (IPS) has been ongoing since the last two decades. There are several suggestions found in literature to fulfill this gap, such as: a) GPS pseudolites (repeaters), b) GPS high-sensitivity receivers working at very low SNR levels, c) the use of cellular signals such as 4G LTE, d) other signals such as RFID, Bluetooth, Lidar, etc., and finally e) wireless local area network (WLAN), commonly known as Wi-Fi. Due to its wide deployment and popularity, WLAN is recently adopted for IPS. There are many WLAN-based IPS methods proposed in literature, which typically use the received signal strength (RSS) measurement available from most WLAN cards at 1 Hz rate. Two main models are proposed in literature for WLANbased IPS: analytical, and empirical. Analytical models use triangulation methods such as angle-of-arrival, or trilateration methods such as time-of-arrival, and time-difference-of-arrival [2]. Other solutions rely on propagation models [3], or require additional hardware for deployment. A drawback from said techniques is a requirement for LOS and/or extra sensors and antennas as well as their analytically complex nature. More recently, empirical-based techniques based on classification are preferred, such as fingerprinting (FP) [4], [5]. The FP-based method uses two stages: offline and online. In the offline stage, RF signals are collected for a discrete grid of locations in an indoor setting to build a database, the radio-map. For each discrete location, or reference point (RP), a signature is generated from the collected RF signals. The radio-map is then used in the online phase for user navigation via classification techniques. Such classification algorithms can be deterministic, i.e., K-nearest-neighbor (KNN) [4], probabilistic such as maximum likelihood estimation (MLE) [5]-[9], and more recently machine learning approaches such as support-vector machines [10], and deep learning [11]-[15]. Most FP-based IPS employ RSS measurements as fingerprints for radio-map construction and online navigation [4]. However, RSS suffers from performance degradations in complex environments due to well-known multipath fading phenomena, adding to accuracy degradation [11]. Additionally, measurement miss-rates are known to degrade overall performance [16]. Recent literature has proposed the channel estimate as an alternative. The channel estimate describes the indoor environment due to multipath fading and other phenomena such as reflections, and refractions. These estimates characterize a specific indoor location with fine-grained information from the fading phenomena [10], [11], [14]. Therefore, the channel state information (CSI) is used as a fingerprint inherent to FP-based IPS, as opposed to less-stable RSS measurements. The channel estimate can be obtained by approximating predefined training sequences or “preambles” contained in received signal frames [14]. As opposed to the RSS measurement, which is reported on any off-the-shelf WLAN network card, the CSI is not directly available from these cards. Recently, a modified firmware on limited vendor cards, e.g., the Intel 5300, called CSI Tool has been made available [17]. The Atheros AR9580 also reported said firmware modifications for CSI measurement availability [18]. These tools leverage a “sounding” mechanism internally found on the 802.11n protocols between an access point (AP) and a receiver to enable a spatial diversity channel. The CSI obtained from the sounding mechanism is specific to OFDM signals, which contain 64 orthogonal subcarriers (SCs), out of which only 52 are used. The CSI obtained from the CSI Tool (and sounding mechanism) reports 30 specific non-symmetric SC channel states from the potential 52 SCs. As an alternative to obtain the CSI for an IPS, this work proposes SDR-Fi, a fast-prototyping software-defined radio (SDR) receiver capable to obtain CSI measurements in a passive mode for IPS. Compared to state-of-the-art CSIbased solutions that use a hacked hardware and limited access to WLAN channel data, the SDR solution has the ability to extract measurements from any portion of the WLAN receiver chain, including the full 52 SC CSI. SDR is defined as having a hardware part where received signals are digitized through a front-end, and a software part where all baseband processing is done. SDR solutions become popular because of providing full control of receiver baseband modules, so the researchers can integrate and test their methods without redesigning all receiver chains. This becomes an advantage for SDR based research for fast prototyping [10]. As for classification, we explore a state-of-the-art feedforward neural network (FFNN) model for the IPS. The described SDR receiver achieves real-time packet collection rate of almost 10 Hz by use of acceleration features from a fast-prototyping software platform [10]. Finally, the proposed FFNN is compared against two existing representative methods: (1) RSS-based Horus [5]; (2) and CSIbased DeepFi [11]. SDR-Fi is examined in representative testing scenarios for both indoor (cluttered), and outdoor (less cluttered) environments for a single AP.