Applying machine learning approaches to indoor positioning

Abstract

For commercial demands, common use cases such as shopping in malls, or supermarkets, warehouse management, game development and so on, are continuously improved. This is where positioning systems, or more specifically indoor positioning systems play an important role. For outdoor environmentws the Global Positioning System (GPS) is nowadays the most relyable technology, but indoor location and navigation remain unresolved issues, and alternative methods are needed. In this scenario the Wi-Fi positioning system is a pretty widely used technology, indeed since large buildings mostly already have a high number of Wi-Fi routers distributed, it makes sense to use the Wi-Fi signals for positioning. Namely, it is possible to employ Received Signal Strength (RSS) signals from any 802.11 device, while Round-Trip Time (RTT) signals from any 802.11mc device. This study investigates the implementation of a 802.11-based indoor positioning system that will follow a fingerprinting approach: the Wi-Fi signal collected at a given point will be compared with a map of measurements previously recorded at several reference points and, by means of machine learning, the most probable location will be extracted from the database. The objective will be to study the real potential of a Wi-Fi based indoor positioning system that simultaneously exploits RTT and RSS information from IEEE 802.11 devices. In order to build the required database, a data collection campaign has been implemented in one of the buildings of Universitat Politècnica de Catalunya. Subsequently, two machine learning algorithms have been implemented using Python¿s scikit-learn and xgboost libraries, and the related predictive performances have been tested. The positioning errors referring to the use of different datasets (e.g., RTT dataset, RSS dataset, the combined RTT/RSS dataset) have been estimated. Next, the results related to positional errors will be outlined, together with the relative benefits, if any, of using a coupled dataset for both considered classifiers.