Sistema de gestão para localização indoor utilizando Wifi Fingerprints e Machine Learning

Abstract

Software solutions that are dependent on the Global Positioning System (GPS) signal may not display good accuracy in indoor spaces (malls, airports, commercial complexes, etc.), as the signal variations are likely to make it unreliable. This work aims to implement an indoor location management solution using signals from local wireless networks to predict the position of a user. To do so, the solution uses WiFi fingerprints, captured by mobile devices with Android operating system, to build bases of instances representing real positions of a user collected in one of the Federal Rural University of Pernambuco - Academic Unit of Garanhuns (UFRPE - UAG) buildings. The mobile application determines the user’s position and a product location management web tool allows the insertion of user points of interest on the indoor map. To complete the process, the mobile application plots a vector from the predicted position of the user to the position of a specific target object, thus composing a complete and functional solution of indoor routes. The algorithms used to predict the position of the user were random forest, multi layer perceptron and Adaboost. The best indoor localization result was obtained with Adaboost, presenting an average error just under 1 meter from the actual position and 98.64% of the results were under an acceptable margin of error (up to 2 meters).