A Sequential Detection Approach to Indoor Positioning Using RSS-Based Fingerprinting



Etemadyrad, Negar

Journal Title

Journal ISSN

Volume Title



Indoor positioning has received significant research and commercial attention during the past two decades. Real world applications include navigating in hospitals, office buildings, warehouses, and parking garages. In this work, we focus on localizing a robot navigating an unfamiliar building. The proposed approach uses received signal strength (RSS) fingerprinting from nearby Wi-Fi access points (APs). RSS fingerprinting approaches are common in indoor localization, and their popularity is in part due to the relatively low hardware cost, taking advantage of the equipment that generates Wi-Fi signals in a wireless local area network (LAN). Moreover, fingerprinting-based localization has significantly lower compu- tational complexity than competing highly sophisticated mathematical techniques, such as probabilistic methods, compressive sensing positioning, and Hidden Markov Models. This thesis presents a sequential detection approach to RSS-based positioning that em- ploys a Bayesian metric to identify the most likely path traveled by the indoor user, given a time series of RSS measurements. The RSS measurements are collected as the robot travels and are passed as input to the proposed algorithm to identify the best user path estimate. A trellis is used to model different possible paths the user could travel, and the Viterbi algorithm is applied to find the most likely path. As a contribution of this work, a sequential metric is developed which takes advantage of Bayes' rule in combination with the k-nearest neighbors (KNN) algorithm to approximate the likelihoods, conditioned on observed RSS, of the paths in the trellis. The performance of the proposed indoor localization algorithm is evaluated using data collected in the Nguyen Engineering building at George Mason University. Data was col- lected at varying times of day and across several days. Training (offline) RSS data was collected at a grid of location points and stored in a database for use during the test (on- line) phase. Various combinations of training and test points were drawn randomly from the collected data and provided as input to the KNN algorithm for estimating the condi- tional likelihood of observations given location. Results show that the sequential detection approach achieves strong performance even when only a small series of RSS measurements are available. In fact, in most cases, there is great improvement in performance, reported both in terms of average distance error and probability of correct path estimation, when a sequence of two sets of RSS measurements is used rather than collecting measurements at only a single location.



Indoor positioning, Localization, RSS fingerprinting, K-nearest neighbors