WiFi Fingerprinting and Tracking using Neural Networks

Abstract

As technology becomes increasingly interconnected through the advent of advanced telecommunications systems such as 5G networking, the necessity for seamless solutions in location-based services becomes increasingly relevant. In response to this demand, there has been a steady increase in academic and commercial interest in indoor real-time locating systems (RTLSs) and in indoor tracking and positioning. With respect to the broader research domain of Positioning, Navigation and Timing (PNT), indoor positioning and tracking offers a number of difficult challenges. Use of the widely diffused Global Navigation Satellite System (GNSS) technology, one of the most accurate sources of position information when it is available, is often infeasible in indoor or obstructed environments [4]. Instead, alternative systems have to be adopted. One approach to positioning and tracking in such environments is fingerprinting, also referred to as mapping or scene analysis. The basic idea of fingerprinting is to build a database containing a collection of measured features at designated reference locations within the environment, and to then perform positioning by applying regression techniques to match new measurements to one or more of those in the database. The position in feature space associated through the database with a particular physical reference location is referred to as the “fingerprint” of the environment at that location, and the assumption that these feature vectors are relatively unique forms the basis for the fingerprinting technique. The fingerprinting procedure typically operates in two stages: an offline stage in which the environment is surveyed at known locations and the results are recorded into a database, and an online stage in which navigation is performed by matching new measurements with the content of the database. Once a match is made, position may be inferred based on the reference positions associated with those database measurements. Fingerprinting offers several important advantages as an indoor positioning technology. One important advantage is that they do not require a specific measurement model: since indoor environments are often segmented and highly non homogeneous in composition, useful observables such as sound and electromagnetic radiation often exhibit highly nonuniform or nonlinear diffuse behaviors which can not be easily or accurately modeled even without advanced and thorough knowledge of the contents of the environment. Fingerprinting replaces the need for an accurate measurement model with a need for an offline training stage, a requirement which may be more realistic in certain contexts. Other advantages are that they are strictly passive and can often make use of existing features of the environment such as installed WiFi networking. For this reason among others, WiFi received signal strength (RSS) fingerprinting is now widely used in indoor positioning and navigation problems [13, 18]. Another remarkable feature of fingerprinting methods is that they do not require any knowledge regarding the location of the transmitting nodes, which is particularly relevant in other schemes such as those which are geometric-based [4]. This contribution presents a novel methodology for performing tracking using RSS fingerprinting. There are two main components in the proposed methodology: i) a filtering technique for tracking the receiver given RSS measurements, which is achieved through a Kalman filter algorithm; and ii) a neural network based data-driven approach for learn the spatial model for the RSS of the field, which is required in the tracking part if a physics-based model is not used. In this paper we particularly focus on the issues related to RSS field changes, that is when the RSS model is learnt by the neural network and used for tracking the receiver but it changes after some time (e.g. due to people walking in the area, obstacles or walls being added, or other situations causing a change in the RSS. To that aim we compare offline training of the neural network (i.e., where a potentially large training data set is recorded and used for training the model) with online training (i.e., where new data is measured at known locations but is more scarce or not available at the same time so it needs to be processed on a sample-per-sample basis) schemes. In this context, we compare a classical neural network training method (i.e., Adam) and a method based on Kalman filtering, which is suited for sequential inference. To validate the proposed RSS field estimation approaches we simulated an indoor environment using the standard path-loss model and trained offline the model, then generated several disturbances in the model and adapt the neural network to incorporate the new data. Results are provided in terms of training error, tracking error, and tracking convergence rate.